e-ISSN 2150-5551

CN 31-2103/TB

REVIEW [Link]

Non‑Invasive Brain‑Computer Interfaces:

Converging Frontiers in Neural Signal Decoding
Cite as
Nano-Micro Lett. and Flexible Bioelectronics Integration
(2026) 18:193

Received: 2 September 2025 Sheng Wang1 *, Xiaobin Song2, Xiaopan Song3 *, Yang Gu3, Zhuangzhuang Cong4,
Accepted: 2 December 2025 Yi Shen2,4 *, Linwei Yu1,3 *
© The Author(s) 2026

HIGHLIGHTS
• The latest advancements in neural signal decoding and the integration of flexible bioelectronics for non-invasive brain-computer
interfaces are reviewed.

• Multimodal data fusion, hardware-software co-optimization, and closed-loop control strategies are critical for enhancing the robust-
ness, adaptability, and real-time performance of brain-computer interface (BCI) systems.

• The robust real-world deployment of BCIs requires breakthroughs in cross-subject generalization, environmental adaptability, and
system reproducibility.

ABSTRACT The development of non-invasive brain-computer interfaces (BCIs)

relies on multidisciplinary integration across neuroscience, artificial intelligence,
flexible electronics, and systems engineering. Recent advances in deep learning have
significantly improved the accuracy and robustness of neural signal decoding. Parallel
progress in electrode design—particularly through the use of flexible and stretch-
able materials like nanostructured conductors and novel fabrication strategies—has
enhanced wearability and operational stability. Nevertheless, key challenges persist,
including individual variability, biocompatibility limitations, and susceptibility to
interference in complex environments. Further validation and optimization are needed
to address gaps in generalization capability, long-term reliability, and real-world opera-
tional robustness. This review systematically examines the representative progress in
neural decoding algorithms and flexible bioelectronic platforms over the past decade,
highlighting key design principles, material innovations, and integration strategies
that are poised to advance non-invasive BCI capabilities. It also discusses the impor-
tance of multimodal data fusion, hardware-software co-optimization, and closed-loop
control strategies. Furthermore, the review discusses the application potential and associated engineering challenges of this technology in clinical
rehabilitation and industrial translation, aiming to provide a reference for advancing non-invasive BCIs toward practical and scalable deployment.

KEYWORDS Non-invasive BCIs; Deep learning; Neural signal decoding; Nanowires; Flexible bioelectronics

Sheng Wang and Xiaobin Song contributed equally to this work.

* Sheng Wang, shengwang_njupt@[Link]; Xiaopan Song, songxiaopan@[Link]; Yi Shen, dryishen@[Link]; Linwei Yu,
yulinwei@[Link]
1
College of Electronic and Optical Engineering & College of Flexible Electronics (Future Technology), Nanjing University of Posts
and Telecommunications, Nanjing 210023, People’s Republic of China
2
Department of Cardiothoracic Surgery, Jinling Clinical Medical College, Nanjing University of Chinese Medicine, Nanjing 210007,
People’s Republic of China
3
School of Electronics Science and Engineering, Nanjing University, Nanjing 210023, People’s Republic of China
4
Department of Cardiothoracic Surgery, Jinling Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing 210007,
People’s Republic of China

Vol.:(0123456789)
 193 Page 2 of 49 Nano-Micro Lett. (2026) 18:193

1 Introduction artifacts and environmental electromagnetic interference, which

result in a low SNR in real-world environments, thereby low-
The development of brain-computer interfaces (BCIs) rep- ering signal quality and affecting the reliability of subsequent
resents a significant integration of neuroscience, artificial analysis [34]. Other non-invasive neuroimaging technologies
intelligence (AI), and bioelectronics, establishing direct com- face similar challenges. For instance, functional near-infrared
munication pathways between neural activity and external spectroscopy (fNIRS) and functional magnetic resonance
technological systems [1–6]. This integration enables effective imaging (fMRI) are constrained by low temporal resolution. In
control of computer interfaces, assistive devices, and robotic contrast, magnetoencephalography (MEG), despite its higher
platforms, thereby enhancing human-technology interactions spatiotemporal resolution, requires strict environmental condi-
[7–11]. Specifically, BCIs are broadly categorized into two tions [35–38]. Furthermore, conventional decoding paradigms,
types: invasive and non-invasive. Invasive BCIs, which utilize including motor imagery (MI), steady-state visual evoked
direct electrode-neural tissue contact, typically achieve high potentials (SSVEP), and P300 event-related potentials (ERP),
signal-to-noise ratio (SNR), sensitivity, and resolution, mak- are constrained by factors such as dependence on user state
ing them suitable for precision interventions requiring high and individual variability. These limitations impede their abil-
signal quality. However, their application is constrained by ity to achieve high robustness in real-world scenarios [39–42].
surgical risks, biocompatibility issues, and long-term stabil- Therefore, the development of non-invasive neural recording
ity challenges. At present, invasive BCIs are mainly used in technologies with millisecond-scale temporal resolution and
clinical settings for advanced cognitive research and motor high SNR remains a key direction for advancing the practical
function restoration in paralyzed patients, which may limit application of high-performance BCIs [43].
near-term direct translation to broader populations, including To support more reliable performance in real-world
healthy individuals [12–21]. In contrast, non-invasive BCIs, applications, recent research is increasingly focused on two
which do not require surgery and offer high safety, have dem- directions: advancing neural signal decoding through AI
onstrated broad scalability and applicability in clinical prac- and integrating flexible bioelectronic platforms. The rapid
tice. These systems are applied in various domains, such as progress of deep learning architectures [44–46], especially
communication, rehabilitation, virtual reality (VR) and aug- convolutional neural networks (CNNs), multimodal hybrid
mented reality (AR), cognitive impairment, mental health, networks, and Transformer models, has improved decoding
fatigue and vigilance monitoring, chronic inflammation, and accuracy and system operational stability, with performance
autonomic nerve regulation [22–24]. Clinically, they serve surpassing that of traditional machine learning methods
as valuable non-pharmacological interventions for conditions [47–50]. Meanwhile, multimodal technological strategies
like amyotrophic lateral sclerosis (ALS) and stroke rehabilita- are continuously expanding the boundaries of non-invasive
tion, contributing to improvements in patients’ quality of life neural monitoring, offering new insights for analyzing spa-
[5, 25, 26]. For instance, Sellers et al. [27] proposed a non- tiotemporal neural dynamics and their regulatory mecha-
invasive BCI system that restored essential communicative nisms [51]. For instance, multimodal stimulation paradigms
functions in ALS patients, enabling reliable letter selection that integrate visual and auditory cues have been shown to
and spontaneous communication in over two-thirds of trials. improve the accuracy and robustness of neural signal decod-
Similarly, Biasiucci et al. [26] integrated a non-invasive BCI ing. In summary, these technological trends indicate feasi-
with functional electrical stimulation, showing improvements ble pathways toward high-performance non-invasive BCIs,
among chronic stroke patients. Beyond its medical applica- although related engineering implementation and clinical
tions, BCI technology is being increasingly applied in enter- translation challenges still need to be addressed [52].
tainment and automotive safety domains, highlighting its Traditional rigid electrodes, due to their high mechanical
broad societal impact [28–31]. stiffness and limited ability to conform to the skin surface,
Electroencephalography (EEG) remains the primary signal frequently cause discomfort and unstable contact. This can
acquisition method for non-invasive BCIs, valued for its non- lead to a degradation in neural signal quality over time, poten-
invasiveness, high temporal resolution, and clinical applicability tially affecting the accuracy and reliability of BCIs [53–55].
[32, 33]. However, EEG signals are susceptible to physiological Recent advances in flexible bioelectronics provide promising

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 3 of 49 193

Fig. 1  Non-invasive BCI: converging frontiers in neural signal decoding and flexible bioelectronics integration

solutions for enhancing the electrode-skin interface [56–63]. or hydrogels hold promise for reducing electrode-skin inter-
By employing flexible polymer substrates and stretchable face impedance, mitigating motion artifacts, and improving the
conductive materials, electrodes have shown notable improve- SNR [69–77]. Looking ahead, combining these flexible sens-
ments in mechanical compliance, conductivity, and resistance ing platforms with advanced deep learning decoding methods
to deformation [53, 64–66], thereby enhancing wearing com- is expected to further improve the performance of non-invasive
fort and signal stability [62, 67, 68]. In particular, electrode BCIs further. Specifically, the innovation in flexible electrodes
designs based on flexible conductive films, nanowire materials, establishes the hardware foundation for high-performance
 193 Page 4 of 49 Nano-Micro Lett. (2026) 18:193

BCIs. By improving interfacial contact and suppressing motion tissue layers (e.g., the skull and cerebrospinal fluid), it may
artifacts, these electrodes provide more stable, high-SNR raw cause spatial blurring and amplitude attenuation of the sig-
signals, thereby supplying high-fidelity data for subsequent nals [81–83]. Consequently, scalp-recorded EEG typically
decoding. Building on this platform, advanced algorithms, exhibits low spatial resolution, low SNR, and high suscep-
such as deep learning, leverage their powerful feature learn- tibility to physiological artifacts and environmental noise.
ing capabilities not only to achieve significant improvements Additionally, individual variations in anatomical structure,
in decoding performance compared to traditional methods but physiological states (e.g., fatigue, attention), and psychologi-
also to actively compensate for inherent hardware limitations cal states (e.g., motivation, emotion) further exacerbate the
such as residual noise and cross-subject variability. This syn- non-stationarity and time-varying nature of neural signals
ergistic paradigm, which embodies the principle of hardware [84–86]. These challenges not only undermine the generali-
laying the foundation and algorithms driving advancement, zation of decoding algorithms but also hinder the stable and
forms a positive feedback loop that collectively expands the reliable decoding of user intent [87, 88].
performance boundaries of non-invasive BCIs [78–80]. To address these challenges, a typical neural signal decod-
This review provides a comprehensive overview of the latest ing pipeline generally comprises three key stages: preproc-
advancements in neural signal decoding, flexible bioelectron- essing, feature extraction, and classification. In recent years,
ics, and their synergistic integration in the field of non-inva- with the rapid advancement of deep learning technologies,
sive BCIs, highlighting their impact on clinical and industrial BCIs have shown improvements in overall performance
applications. The article is structured as follows: Following [89–91]. A notable trend in current research is the shift from
the introduction in Sects. 1 and 2 reviews pivotal advances in "open-loop, static" systems to "closed-loop, adaptive" para-
neural signal decoding, including multimodal stimulation para- digms [92], with development direction focusing on enhanc-
digms, multimodal neural signal acquisition technologies, and ing real-time decoding capabilities, improving asynchronous
the significance of dynamic neural decoding and closed-loop detection mechanisms, and optimizing shared control strate-
control strategies driven by deep learning. Section 3 critically gies. The core driving force behind this transformation is AI-
explores technological innovations in flexible bioelectronics, driven closed-loop human-machine interaction architectures.
including interface optimization mechanisms for conduc- These architectures not only integrate multimodal sensing
tive thin films, miniaturization strategies for nanowire-based and deep learning models but also enable dynamic coordi-
devices, multiphysics coupled design of wearable systems, and nation and bidirectional adaptation between the brain and
hardware-software co-optimization with deep learning archi- external devices through a "perceive-decode-applications"
tectures. Section 4 addresses current technological challenges closed-loop mechanism [88].
and future opportunities, focusing on multimodal neural sens-
ing, the enhancement of robustness in adaptive closed-loop
systems, and the development of scalable clinical translation 2.1 Multimodal Stimulation Paradigms for Enhanced
pathways. Through a comprehensive analysis of interdisci- Neural Decoding
plinary technological convergence, this review establishes a
tripartite framework (decoding-sensing-application) for next- Research in non-invasive BCIs is increasingly focusing on the
generation non-invasive BCIs and provides actionable insights integration of multimodal paradigms to significantly enhance
to support their integration into precision medicine and intel- overall system performance. This performance improvement
ligent human-machine interaction scenarios (Fig. 1). is primarily achieved through advanced multimodal fusion
algorithms that effectively integrate information from dif-
2 Advancing Neural Signal Decoding ferent sensory channels such as visual, auditory, and tactile,
Methodologies thereby optimizing the accuracy and robustness of neural sig-
nal decoding [93, 94]. Specifically, multimodal information
The core performance limitations of non-invasive BCIs processing first requires the construction of efficient fusion
primarily originate from inherent biophysical constraints. algorithms, such as signal analysis based on manifold geom-
When cortical neural activity passes through multiple etry, feature extraction using common spatial patterns (CSP),
and signal calibration leveraging source imaging priors

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 5 of 49 193

Fig. 2  Schematic illustration of multimodal stimulation paradigms for neural signal decoding. a Left: An example trial of the flash-beep para-
digm. Right: Across-participants’ mean illustrating absolute differences. Reproduced with permission [52]. Copyright 2019, Springer Nature.
b Left: Amplitude distribution of evoked EEG signals in response to the words "one" to "nine". Right: The language and frequency spectra
of the feedback EEG signals in MASSR-EEG. Reproduced with permission [98]. Copyright 2023, Wiley-VCH GmbH. c System performance
under different stimulus conditions and pairwise classification accuracy. Reproduced with permission [99]. Copyright 2016, World Scientific
Publishing Company. d Stimulation interface for the P300 and SSVEP paradigms. Reproduced with permission [100]. Copyright 2015, IEEE. e
Framework of the hybrid EEG signal processing module combining SSVEP and P300. Reproduced with permission [41]. Copyright 2022, Else-
vier. f Experimental designs for the three BCI paradigms. Reproduced with permission [101]. Copyright 2019, Oxford University Press. g Mul-
tidimensional paradigm combining perception, sustained attention, selective attention, and episodic memory tasks, with transfer learning used to
predict memory performance. Reproduced with permission [102]. Copyright 2019, Springer Nature. h Visual representation of emotion-related
components across cortical visual areas. Reproduced with permission [103]. Copyright 2019, The American Association for the Advancement of
Science

[95–97]. These methods represent preliminary approaches was further demonstrated that pre-stimulus neural oscilla-
to multimodal integration and modestly improve decod- tions (such as alpha and gamma power) modulate the causal
ing accuracy. Rohe et al. [52] combined EEG with Bayes- "prior", thereby elucidating the neural basis of perceptual
ian modeling to reveal the neural dynamics of hierarchical decision-making with temporal precision. Methodologically,
Bayesian causal inference in multisensory perception. The the original authors employed rigorous statistical approaches,
work revealed that the brain does not simply integrate sen- including cluster-based multiple comparison correction and
sory signals, but dynamically arbitrates between integration bootstrapped confidence intervals, to substantiate their con-
and segregation strategies based on intersensory conflict, clusions (Fig. 2a). Notably, Li et al. [98] developed a flexible
computing the final percept through "model averaging". It electrode, which combined SSVEP with multimodal auditory
 193 Page 6 of 49 Nano-Micro Lett. (2026) 18:193

steady-state responses (MASSR). The bimodal paradigm sig- scores (4.50 ± 0.51 vs. 2.85 ± 0.67), indicating enhanced
nificantly outperformed the unimodal conditions, achieving operational efficiency and improved user experience.
an average recognition accuracy of 89.6% in the occipital Although the means and standard deviations reported by
channels (O1, Oz, O2), while the MASSR paradigm attained the original authors provide preliminary evidence of the
only 36.7% in the temporal channels (FT9, FT10, TP9, system’s advantages, more comprehensive statistical analy-
TP10). When the SSVEP-MASSR multimodal paradigm was ses in future studies will be essential to rigorously validate
employed, Li et al. observed that the accuracy improved to these performance claims (Fig. 2e). Crucially, multimodal
90.4% in the visual channels and 54.0% in the auditory chan- integration exhibits neuroplasticity across both spatial and
nels. This work served as a proof of concept for the paradigm temporal dimensions, further enhancing the robustness
using the nine words ("one" to "nine"). Extending its appli- of neural decoding. Zhang et al. [39] designed a hybrid
cability to more complex semantic scenarios is a key next non-invasive BCI that integrates MI and high-frequency
step for further development (Fig. 2b). Beyond audiovisual SSVEP (34/35 Hz) to control a wearable soft robotic glove
integration, the incorporation of the tactile modality further designed for stroke rehabilitation. Methodologically, they
expands the potential of multimodal BCIs. Yin et al. pro- employed the filter bank common spatial pattern (FBCSP)
posed an innovative auditory-tactile bimodal P300 BCI [99]. to process MI signals and the filter bank canonical cor-
The results showed that the bimodal paradigm significantly relation analysis (FBCCA) to decode SSVEP signals,
outperformed the unimodal conditions, achieving an online ultimately making decisions through a weighted fusion
information transfer rate (ITR) of 10.77 bits ­min−1, which algorithm. Online experimental results demonstrated the
represented an improvement of 45.43% (p < 0.05) and 51.05% system’s feasibility, achieving mean classification accu-
(p < 0.001) over the auditory-only and tactile-only paradigms, racies of 95.83% ± 6.83% in 12 healthy subjects and
respectively. It also achieved a higher average accuracy of 63.33% ± 10.38% in 9 stroke patients. This high-frequency
88.67% with fewer trials (average 2.92). Future work explor- SSVEP reduced visual fatigue, while decomposed action
ing its application to more complex scenarios is a promising imagery improved task intuitiveness, forming a closed-loop
and anticipated direction for development (Fig. 2c). "peripheral-central-peripheral" rehabilitation framework.
Fan et al. [100] developed a hybrid BCI system based Meanwhile, although the dataset study by Lee et al. pro-
on P300 and SSVEP for vehicle destination selection. The vided a theoretical foundation for the necessity of multi-
system achieved an average accuracy of 98.93% ± 0.48% and modal BCIs [101]. The study revealed that the MI para-
a mean selection time of 25.95 ± 1.04 s under real driving digm had the highest rate of BCI illiteracy (53.7%), while
conditions. The study represents an important step forward the exogenous paradigms—ERP and SSVEP—exhibited
by validating the system’s performance in real driving envi- much lower illiteracy rates of approximately 10%. More
ronments with fluctuations in illumination and noise. Future importantly, all participants were able to effectively con-
work could systematically evaluate its performance under a trol at least one paradigm, with no user being classified as
broader spectrum of complex and extreme conditions to fully universally illiterate. These results suggest that by devel-
define its operational limits (Fig. 2d). oping adaptive systems that integrate multiple paradigms,
Tang et al. [41] developed a BCI painting system that it is possible to leverage the complementary strengths of
employed a hybrid control approach, combining SSVEP and different modalities, thereby extending control to a wider
P300. The system achieved an average tool selection accu- range of users (Fig. 2f).
racy of 88.92% ± 3.94%, with an ITR of 74.20 ± 5.28 bpm In the domain of more complex passive BCIs and
in the copy-painting task and 71.80 ± 5.15 bpm in the advanced cognitive state decoding, multimodal integration
free-painting task. The average tool selection accuracy likewise demonstrates promise. Mirjalili and Duarte devel-
was slightly higher than that of the traditional P300-only oped a transfer learning framework integrating EEG data
system (86.78% ± 4.56%), while the ITR was signifi- from sustained attention [102], selective attention, and visual
cantly superior to the conventional system (copy-painting: perception tasks to decode the encoding of episodic memory.
60.86 ± 6.56 bpm; free-painting: 58.93 ± 6.11 bpm). Sub- Their findings suggested that tracking multidimensional cog-
jective evaluations demonstrated significantly higher user nitive states enhanced predictive validity through distributed
satisfaction (4.10 ± 0.64 vs. 3.40 ± 0.75) and motivation neural engagement patterns, improving memory encoding

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 7 of 49 193

Fig. 3  Schematic illustration of multimodal signal acquisition. a Left: EEG and fNIRS features calculated by TRCA. Right: EEG and fNIRS
channels. Reproduced with permission [107]. Copyright 2022, IEEE. b EEG-informed fMRI and connectivity analyses in the value-based task.
Reproduced with permission [108]. Copyright 2017, Springer Nature. c Left: 140 ms full-space event topography maps. Right: ITR and clas-
sification accuracy of single-modality EEG, single-modality MEG, and MEG-EEG fusion. Reproduced with permission [109]. Copyright 2024,
IEEE. d Synchronization performance. Reproduced with permission [110]. Copyright 2017, Springer Nature. e Neurovascular coupling mod-
eling. Reproduced with permission [111]. Copyright 2025, John Wiley and Sons

prediction accuracy from 72% to 81.4% (p < 0.001). A more 2.2 Multimodal Signal Acquisition: Techniques
comprehensive understanding of the underlying neural and System Design
mechanisms may be pursued in future work by leveraging
complementary neuroimaging techniques (Fig. 2g). Kragel EEG remains the foundational method for non-invasive
et al. [103] expanded this framework by demonstrating how BCI signal acquisition because of its high temporal reso-
affective states are spatially encoded across visual hierarchy lution and non-invasive characteristics. However, EEG is
regions (V1-PH), enabling decoding of complex cognitive- susceptible to artifacts and has limited spatial resolution,
affective interactions (Fig. 2h). To address the challenges in which has led to growing interest in integrating multimodal
multimodal integration and further enhance the performance neuroimaging techniques. Recent approaches combine EEG
of non-invasive BCIs, innovative deep learning networks with complementary modalities: fNIRS provides hemody-
will be a crucial future direction. Particularly, Transformer namic data, while MEG enhances spatiotemporal resolution.
networks suitable for multimodal signal analysis, high-order This integration effectively addresses the inherent trade-off
feature fusion techniques, and algorithms based on cross- between temporal and spatial resolution in non-invasive sys-
modal alignment will play a pivotal role. tems [104–106].
 193 Page 8 of 49 Nano-Micro Lett. (2026) 18:193

Recent studies in the field of BCIs have demonstrated 16-bit signal transmission link between visual stimuli and
that combining EEG and fNIRS can enhance system per- MEG data acquisition, ensuring precise synchronization
formance. Jiang et al. [107] developed a unified EEG-fNIRS between stimulus onset and data recording initiation. With-
bimodal signal processing framework to analyze single-trial out additional signal-domain preprocessing, the research
neural signals during robot-assisted bimanual cyclical tasks. team directly fused the two modalities and decoded them
They employed the artifact subspace reconstruction algo- using the multiclass discriminative canonical pattern match-
rithm to remove large-amplitude artifacts from EEG, and ing algorithm. Under a 4-s data window, the fusion model
applied the temporal derivative distribution repair method achieved an average classification accuracy of 91.71%,
to correct motion artifacts and baseline drift in fNIRS signals. markedly exceeding that of MEG alone (88.57%) and EEG
The analysis time window was defined from 2 s before to (60.76%), with corresponding p-values below 0.01 and 0.001.
4 s after task execution, during which event-related desyn- This performance gain was attributed to MEG’s heightened
chronization/synchronization (ERD/ERS) and oxygenated sensitivity to contralateral polarity reversal in the occipital
hemoglobin concentration changes (ΔHbO) were extracted cortex, confirming the critical role of multimodal acquisition
as features. In terms of statistical methods, Jiang et al. found in enhancing spatial decoding performance in SDMA-based
statistically significant differences in both ERD/ERS and BCI systems (Fig. 3c). The Syntalos framework, developed
ΔHbO responses among the three bimanual tasks during by Klumpp et al. [110], is designed to address the challenge
specific time intervals (p < 0.001), with post hoc analyses of precise synchronization in multimodal data acquisition,
indicating that the anti-phase task elicited the strongest particularly for long-duration recordings. The framework
activation using statistical analysis. For classification, they established a globally shared master clock and continuously
used a support vector machine, which demonstrated that the performs statistical analysis and correction of timestamps
fused EEG-fNIRS features achieved an accuracy of 90.1% from various devices, ensuring precise alignment of all input
in discriminating the three bimanual movement patterns— timestamps. Experimental validation demonstrated that Syn-
significantly higher than the accuracy using single-modality talos maintains stable synchronization across devices for over
features (EEG: 74.8%; fNIRS: 82.2%) (Fig. 3a). FMRI offers 24 h. In a simulated time-drift experiment (with a systematic
extremely high spatial resolution, enabling precise localiza- drift of 1 ms per second), they observed that the accuracy of
tion of brain activity. However, its temporal resolution is con- the behavioral classifier dropped from nearly 100% to near-
strained by the slow nature of the hemodynamic response. chance levels when the temporal misalignment between
Simultaneous EEG-fMRI acquisition and fusion can combine neural spike signals and whisker-touch behavioral data accu-
the high temporal resolution of EEG with the high spatial mulated to 1 s. This demonstrated that Syntalos addressed
resolution of fMRI, providing comprehensive spatiotemporal the challenge of precise synchronization in multimodal
information about brain activity. This multimodal integration data acquisition, particularly for long-duration recordings
has been successfully applied in various BCI applications, (Fig. 3d). Li et al. [111] proposed the BrainFusion framework
including emotion recognition and MI. Pisauro et al. [108] to address challenges in reproducibility and deployment in
employed a snack choice task combined with simultane- multimodal BCI research. The framework manages signal
ous EEG-fMRI and computational modeling to investigate temporal synchronization through standardized data con-
the role of the posterior medial frontal cortex (pMFC) in tainers and two alignment strategies (time-point alignment
value-based decision-making. To mitigate artifacts inherent and event-based alignment), and supports the BIDS format
to simultaneous recording, the study used twisted leads to to enhance data comparability. Case studies demonstrated
reduce electromagnetic interference and performed offline its efficacy: it achieved 95.5% accuracy in EEG-fNIRS MI
removal of gradient artifacts. Results revealed that bold sig- classification and deployed an EEG-ECG sleep staging
nal activity in the pMFC was significantly associated with model (80.2% accuracy) as an executable end application.
these EEG dynamics, suggesting its role as a neural sub- The framework is primarily designed for offline analysis.
strate for evidence accumulation during value-based choices Extending it to real-time closed-loop control constitutes a
(Fig. 3b). Ji et al. [109] constructed a 16-command SDMA- valuable direction for future research (Fig. 3e). The dual
encoded MEG-EEG fusion modality BCI system. The study capacity to resolve spatiotemporal dynamics and endogenous
implemented a synchronous triggering mechanism via a modulation establishes multimodal approaches as critical for

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 9 of 49 193

clinically viable BCIs [112]. To address multimodal engi- be developed to optimize the efficiency of the processing
neering challenges, Bayesian frameworks and advanced algo- pipeline and reduce computational complexity [114]. Time
rithms have been widely applied in EEG source imaging to jitter and misalignment caused by systems with different
address source localization spatial priors [113]. To minimize sampling rates require coordinated advances in both precise
mutual electromagnetic interference between sensors of dif- hardware synchronization and software-level coordination
ferent modalities, in addition to electromagnetic shielding [115, 116].
techniques, signal processing filters tailored to specific noise
characteristics should be integrated [84]. Moreover, in online
or closed-loop systems, the phase lag inherent in signal trans- 2.3 Single‑Model Deep Learning Architectures
mission, processing, and actuator response cannot be ignored, for Neural Dynamics Interpretation
as it directly determines the timeliness and effectiveness of
closed-loop interventions. To compensate for such delays, Deep learning has become a key approach for decoding non-
advanced signal processing and predictive algorithms must invasive neural signals [117–119], effectively addressing

Fig. 4  Schematic illustration of single-model deep learning architectures. a Comparison of BCI Decoder Performance with Random Levels and
Traditional AR Methods. Reproduced with permission [121]. Copyright 2024, Springer Nature. b Top: Classification performance comparison
of ablation studies on two datasets. Bottom: Classification performance comparison of models using different convolutional layer filter blocks.
Reproduced with permission [123]. Copyright 2022, Elsevier. c Top: Accuracy of five models on the benchmark dataset. Bottom: Classification
accuracy of the model under human intervention. Reproduced with permission [124]. Copyright 2025, Elsevier. d Left: Training loss of the gen-
erator and discriminator, and schematic of the model algorithm. Right: t-SNE visualization. Reproduced with permission [125]. Copyright 2021,
IEEE
 193 Page 10 of 49 Nano-Micro Lett. (2026) 18:193

challenges such as spectral-temporal complexity, individual EEGNet using the Wilcoxon test, with statistical results con-
variability, and data scarcity. Although deep learning has firming the significant advantages of their method. Further-
achieved significant progress in enhancing decoding per- more, the study included an ablation experiment to assess
formance for non-invasive BCIs in recent years, its practi- the functional contributions of different components of the
cal application still faces substantial challenges. These chal- model. To promote research reproducibility, the authors have
lenges represent the critical bottleneck in transitioning BCIs made the relevant code publicly available (Fig. 4b). Ju and
from laboratory settings to real-world deployment. So, an Guan proposed Graph-CSPNet [96], which integrated graph
increasing number of recent studies have begun to actively neural networks with SPD manifolds for MI classification
simulate realistic conditions, introducing controlled pertur- tasks. After establishing its excellent time-frequency feature
bations to probe the performance limits of algorithms and extraction capability and classification effectiveness, they
thereby to reveal their applicability boundaries in real-world further enhanced model interpretability through spectral dis-
scenarios [120]. tribution visualization and graph Laplacian analysis. The
In terms of model architecture, researchers have explored approach clearly demonstrated the model’s attention pat-
diverse approaches centered around the spatiotemporal terns toward different frequency bands in the time-frequency
structure, manifold properties, and spectral patterns of EEG. graph structure, including the θ, μ, β, and γ rhythms. The
Forenzo et al. [121] constructed an EEG-based BCI dataset research team specifically emphasized that the frequency
comprising 28 subjects and 168 h of recordings, designed for bands prioritized by their model show alignment with estab-
online continuous pursuit tasks. Results demonstrated that lished neurophysiological mechanisms of MI. Subsequently,
deep learning (EEGNet, PointNet) decoders significantly they introduced a geometric deep learning framework named
outperformed chance level across all sessions. Forenzo Tensor-CSPNet [97]. This model represented EEG spatial
et al. used a Holm-adjusted Wilcoxon signed-rank test to covariance matrices on symmetric positive definite (SPD)
statistically validate this performance (p < 0.05), thereby manifolds and extracted spatio-spectro-temporal features
supporting the reliability of MI features in the dataset. To of the signals by combining deep neural networks on SPD
systematically evaluate decoding stability under variations manifolds with CNNs. Under cross-session non-stationary
in electrode configuration and physiological noise interfer- scenarios, this model achieved an accuracy improvement of
ence, the authors removed five electrodes and introduced approximately 0.3 compared to FBCSP for Subject No.28
ocular artifacts into the data. Results showed no significant on the MI-KU dataset. Furthermore, the authors conducted
decline in decoding performance (p = 0.463), indicating a visualization analysis of the model’s features using Deep-
the model’s robustness in device variation scenarios. The LIFT and t-SNE. The results showed that the activation pat-
model weights and code have been made publicly available terns captured by the model (such as contralateral activation
(Fig. 4a). Li et al. [122] developed the HR-SNN model to in the C3/C4 regions) were consistent with the neurophysi-
investigate its robustness under noise interference and chan- ological mechanisms of ERD/ERS during MI.
nel loss conditions. When subjected to Gaussian noise with In addition to innovations in network architecture, train-
an amplitude of 10%, the model achieved an accuracy of ing strategies and data augmentation methods also play a
68.59% using subject-specific transfer learning (SSTL). key role in performance improvement. The FBCNN-TKS
Under a more severe scenario with 16 out of 64 channels model proposed by Huang et al. [124] demonstrated strong
randomly removed, the SSTL accuracy remained at 71.48%. performance in offline analysis. With a 0.4-s data length,
In addition, Li et al. [123] proposed a Riemannian convo- the model attained a classification accuracy of 83.10% and
lutional neural network (RMCNN) that integrates a spatial- an ITR of 251.54 bpm on the Benchmark dataset, while
temporal convolutional layer with a Riemannian block. The achieving 72.98% accuracy and an ITR of 203.47 bpm on
model was evaluated for cross-session generalization across the BETA dataset. Huang et al. used a paired t test to validate
three public MI datasets. Experimental results demonstrated the model’s efficacy, with results confirming that FBCNN-
that the model achieved an offline classification accuracy of TKS significantly outperformed traditional machine learning
80.52% on the BCI Competition IV 2a dataset. To validate models as well as deep learning methods such as eTRCA
its effectiveness, Li et al. conducted a systematic comparison and DNN (p < 0.01 on Benchmark, p < 0.001 on BETA).
between RMCNN and baseline methods such as FBCSP and The study systematically validated the effectiveness of each

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 11 of 49 193

model component through rigorous experiments: electrode Fahimi et al. [125] proposed a framework based on con-
configuration tests identified 9 or 32 channels as the optimal ditional deep convolutional generative adversarial networks
topology; ablation studies revealed that removing the fil- (DCGANs) to enhance training data by generating artificial
ter bank, TKS module, or center loss consistently degraded EEG signals, addressing the challenge of degraded classifi-
performance, especially under shorter data lengths; and a cation performance in BCI systems under diverted attention
hyperparameter sweep determined the optimal center loss conditions. The work validated the diversity and realism of
weight λ to be 0.0005. These experiments collectively sup- the generated signals through quantitative metrics, such as
port the validity of the model architecture and the synergistic the closeness of the KL divergence between generated and
interactions among its components (Fig. 4c). real samples, and visualization analyses, including t-SNE
embeddings and spectrogram comparisons. The adversarial

Fig. 5  Schematic illustration of single-model deep learning architectures. a Robustness of deep PNNs against unpredictable external perturba-
tions. Reproduced with permission [127]. Copyright 2023, The American Association for the Advancement of Science. b Transfer learning
algorithms on the EEG dataset. Reproduced with permission [128]. Copyright 2024, Springer Nature. c Left: Comparisons of the Spiking Jelly
with other frameworks and applications. Right: Typical applications of Spiking Jelly. Reproduced with permission [129]. Copyright 2023, The
American Association for the Advancement of Science. d Introducing biological SBP into SNNs. Reproduced with permission [130]. Copyright
2021, The American Association for the Advancement of Science
 193 Page 12 of 49 Nano-Micro Lett. (2026) 18:193

training process demonstrated stability, with generator and neural network (Ray-BNN), demonstrating potential in transfer
discriminator losses converging after approximately 300 learning for dimension-varying tasks. The model preserved
iterations, and no mode collapse was observed. In model trained weights and supported architectural adaptation during
evaluation, the method improved classification accuracy in network expansion through dynamic 3D neuronal connectivity
the diverted attention condition from a baseline of 73.04% to and sparse matrix representations. On an EEG dataset com-
80.36%. To further assess generalization, the research team prising multiple paradigms, Ray-BNN achieved an accuracy
tested the model on the BCI Competition III Dataset IVa, of 85.6% (p ≤ 1.7968 × 10⁻3) in 54-fold subject-independent
where classification accuracy for different MI tasks increased testing by integrating feature extractors from Deep4Net and
from 67.57% to 71.14% (p < 0.02). These results collectively X-dawn, indicating robust cross-subject generalization. Dur-
indicated that the proposed data augmentation strategy not ing transfer learning on the Alcala dataset, input dimensions
only effectively enhanced BCI classification performance progressively increased from 6 to 162 access points. The
under diverted attention but also exhibited a degree of cross- model reduced cumulative training time by approximately
dataset generalization capability (Fig. 4d). Zhang et al. [126] 85% compared to BiLSTM while maintaining sparsity below
designed a multiple-source prototype-supervised adversarial 40% in the weight matrix. The study provided publicly avail-
transfer learning approach (PSAT). By constructing a multi- able code and data to support reproducibility. The foundation
source fusion framework that weighted integrates different for cross-subject generalization established by this study pro-
source domains, PSAT reduced cross-subject discrepancies vides a clear direction for future work: verifying the stability
and intra-subject non-stationarity. It also addressed the chal- of transfer performance in scenarios with device heterogene-
lenges of cross-subject EEG classification in BCI. To rig- ity will drive substantial advancements of the framework in
orously validate the effectiveness of their method, Zhang complex application environments (Fig. 5b). Spiking neural
et al. implemented a linear step-up procedure to control the networks (SNNs), as pivotal enablers of brain-inspired intel-
false discovery rate, thereby mitigating the risk of false posi- ligence, have recently seen significant advances in both soft-
tives in multiple comparisons. Experimental results on three ware tools and learning algorithms. At the framework level,
MI-EEG datasets demonstrated that PSAT achieved higher the Spiking Jelly framework introduced by Fang et al. [129]
classification accuracy than its variants, PSAT-c, PSAT-a, has effectively bridged a critical gap in the dedicated software
and PSAT-v. Further ablation studies revealed that both the toolchain for spiking deep learning. Spiking Jelly significantly
prototype mapper and domain discriminator were indispen- reduced the technical barrier to SNN research and promotes
sable modules for enhancing system performance. the development of the software ecosystem for neuromorphic
In addition to the aforementioned models, Momeni et al. computing (Fig. 5c). At the algorithmic level, the biologically
[127] proposed the physical local learning (PhyLL) algorithm, inspired self-backpropagation (SBP) mechanism proposed by
establishing a paradigm for training physical neural networks. Zhang et al. [130] facilitates coordinated weight adjustment
To evaluate the robustness of the system, the original authors in both SNNs and ANNs by allowing synaptic modifications
introduced Gaussian noise perturbations (mean μ = 0.1–0.5, (potentiation or depression) at output neurons to backpropa-
standard deviation σ = 0.25–0.50) into the optical system. gate across layers to upstream synapses (Fig. 5d). Current
Under severe perturbations caused performance degradation, research on deep learning-based decoding for non-invasive
PhyLL recovered high accuracy within just a few training BCIs is evolving from singular performance optimization
epochs. In contrast, the digital model-dependent physics- toward building multidimensional capabilities. Architectur-
aware backpropagation (PA-BP) method exhibited its accuracy ally, models are becoming more aligned with the spatiotem-
decline to approximately 55% even under mild perturbations poral and geometric properties of neural signals. In training,
and failed to recover. After validation on multiple physical methods increasingly integrate data augmentation, transfer
platforms, this work demonstrated the adaptive capability learning, and alignment strategies to enhance generalization.
of physical neural networks in unstable environments and Nevertheless, there remains a notable absence of exploration
laid the groundwork for future exploration of their potential into performance boundaries under real-world conditions such
in BCIs to address challenges such as channel characteristic as channel loss, noise interference, and attentional fluctuations.
variations and environmental interference (Fig. 5a). Yuen Most approaches remain focused on offline analysis, and the
et al. [128] designed a three-dimensional ray-traced biological generalization limits in dynamic environments and with device

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 13 of 49 193

Fig. 6  Hybrid deep learning frameworks for multiscale signal decoding. a Architecture of the CNN-LSTM. Reproduced with permission [131].
Copyright 2025, Elsevier. b Architecture of CNN-LSTM. Reproduced with permission [132]. Copyright 2022, Elsevier. c Architecture of CNN-
BiLSTM network and recognition accuracy for different methods. Reproduced with permission [133]. Copyright 2024, Elsevier. d Decoding
process of deep neurolinguistic learning in real time. Reproduced with permission [134]. Copyright 2023, IEEE. e Multimodal deep fusion
framework and hyperparameter tuning using Bayesian optimization. Reproduced with permission [135]. Copyright 2024, IEEE

heterogeneity require more systematic evaluation, which rep- In the area of motor decoding and electromyographic
resents a critical direction for future research. activity prediction, existing methods often struggle to cap-
ture the nonlinear relationship between EEG and EMG sig-
nals. Therefore, developing nonlinear models capable of
2.4 Hybrid Deep Learning Frameworks for Multiscale
integrating spatiotemporal features, such as the CNN-LSTM
Signal Decoding
model, is crucial for extracting muscle activity-related
information. Amiri and Shalchyan proposed a deep learn-
Single-model decoding approaches, though prevalent, are
ing model integrating CNN and LSTM networks to decode
often limited by their generalizability, especially when faced
muscle activity from non-invasive EEG signals [131]. Dur-
with complex, real-world data. In contrast, hybrid architec-
ing the grasp and lift (GAL) task, their model achieved an
tures have emerged as a promising solution by integrat-
average correlation coefficient (CC) of 0.76 ± 0.10 and a
ing spatial, temporal, and spectral representations. These
normalized root mean square error (nRMSE) of 0.21 ± 0.05
approaches effectively address key challenges such as cross-
between the actual and predicted muscle activities for two
trial variability and signal complexity, setting the stage for
muscles. They employed statistical tests with Bonferroni
robust decoding in both motor and linguistic domains.
 193 Page 14 of 49 Nano-Micro Lett. (2026) 18:193

correction (adjusted p-value < 0.016) confirmed that the arm. The study designed multiuser collaborative scenarios,
model significantly outperformed two comparative methods: allowing different users to control the prosthetic arm for
multivariate linear regression (mLR) and multilayer percep- themselves or their partners. The decoded neural language
tron (MLP). Building on these promising results, several is used to perform high-level tasks such as object delivery
aspects emerge as valuable avenues for further investigation. and emotional interaction. The overall average task success
For instance, while the current evaluation was conducted rate was 72.36%, with an average latency of 4.450 s per word
offline, exploring real-time implementation represents a and 8.645 s for sentence generation (Fig. 6d).
natural and important next step, particularly considering the In the domain of multimodal fusion and electrophysiologi-
computational profile of the model. The average decoding cal source imaging. Jiao et al. [135] proposed a multimodal
time for the CNN-LSTM architecture was 339.76 ± 51.71 s. deep fusion framework using attention neural networks
This provided a useful baseline for future work aimed at (MMDF-ANN). This framework aimed to improve locali-
optimizing efficiency for real-time applications, especially zation accuracy and stability under conditions of extended
when compared to the faster mLR (3.93 ± 3.02 s) and MLP sources and low SNR in electrophysiological source imag-
(95.40 ± 22.44 s) models (Fig. 6a). Although CNN-LSTM ing (ESI). The framework used dual-stream CNN modules
can automatically extract spatiotemporal features, it requires to process EEG and MEG signals separately and integrated
large amounts of data for training. To address the issue of the features through a channel-wise attention mechanism.
data scarcity, Khademi et al. [132] proposed a hybrid CNN- On synthetic data, the model demonstrated strong perfor-
LSTM deep learning model based on transfer learning and mance. For instance, in single-source localization tasks with
data augmentation, aiming to enhance the classification per- an SNR of 30 dB, it achieved an area under the precision-
formance of MI-EEG signals. Their approach employed the recall curve (AUPRC) of 0.944 ± 0.104, outperforming tra-
continuous wavelet transform (CWT) to convert EEG sig- ditional methods such as MNE and sLORETA, as well as the
nals into time-frequency images. They expanded the dataset unimodal deep learning model ConvDip. Ablation studies
fivefold (from 288 to 1440 trials per subject) using a non- confirmed that both multimodal fusion and dilated convolu-
overlapping sliding window cropping strategy. Evaluated tion contributed to the performance improvement. Notably,
on the BCI Competition IV dataset 2a, this model achieved as the number of sources increased, the model’s performance
a mean classification accuracy of 86% and a mean Kappa exhibited a declining trend: in four-source tasks, the AUPRC
value of 81% (Fig. 6b). dropped to 0.827 ± 0.166. This phenomenon suggested that
In the direction of rehabilitation systems. An et al. [133] the diversity of training data may affect the model’s adapt-
developed a real-time classifier based on CNN-BiLSTM for ability to complex activation patterns. In validation with
a patient-centered AR-SSVEP active rehabilitation exoskele- real data, MMDF-ANN successfully localized visual cortex
ton system. The study achieved an offline classification accu- activity in a face perception dataset containing simultane-
racy of 98.5% and an ITR of 210 bits m ­ in−1 within a 0.5-s ous EEG/MEG recordings. In an epilepsy dataset with only
time window. In the online experiments, the subjects took an EEG recordings, the model still achieved more focal source
average of 319 s to complete four non-repetitive trajectory localization compared to ConvDip. These findings establish
tasks, which was 20% longer than the standard reference a solid foundation for multimodal fusion approaches, while
time. As an initial investigation involving six healthy young their broad applicability in diverse clinical environments rep-
participants, this study established a foundation for subse- resents a promising direction for future validation (Fig. 6e).
quent research aiming to expand subject diversity (Fig. 6c). Future work could adopt domain adaptation methods to
In the field of neural language decoding and interactive improve device compatibility and establish comprehen-
control. Jeong et al. [134] proposed a real-time, non-invasive sive performance degradation curves to better define the
neural language decoding method based on deep neurolin- algorithm’s operational boundaries. Together, these meth-
guistic learning for multiuser BCI. This approach combined odological advances illustrate the considerable potential
a CNN and a gated recurrent unit (GRU) to decode speech of hybrid deep learning strategies that integrate multiscale
imagery from EEG signals. It then integrated the decoded feature encoding and cross-modal alignment [136]. While
results into sentence-level neural commands via a rule-based significant gains have been demonstrated across decoding
sentence generation model to control a neural prosthetic tasks and modalities, future work must address cross-subject

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 15 of 49 193

Fig. 7  Schematic illustration of real-time, asynchronous, and collaborative non-invasive BCIs. a Top: Schematic illustration of MI-BCI exper-
imental paradigm. Bottom: Comparative performance of online BCI control. Reproduced with permission [138]. Copyright 2025, IEEE. b
Decoding based on EMG signals. Reproduced with permission [139]. Copyright 2025, Springer Nature. c Left: ErrP decoding results. Right:
Schematic of the control architecture and experimental protocol. Reproduced with permission [140]. Copyright 2021, Springer Nature. d Left:
Schematic of the BCI robot control system and TE of the robot 3D reach task. Right: Comparison of the strategies of direct BCI control and
shared BCI control. Reproduced with permission [142]. Copyright 2023, IEEE

generalizability and robustness in real-world environments FBCSP, in both sessions. The model achieved improvements
[85, 137]. in online task accuracy of 20% and 27%, respectively, and
the main effect of the model was significant (P < 0.0001).
2.5 Toward Real‑Time, Asynchronous, Methodologically, the original authors used Bonferroni cor-
and Collaborative Non‑Invasive BCIs rection for all statistical tests involving multiple compari-
sons, ensuring the rigor of statistical inference. Furthermore,
There is a growing focus in non-invasive BCI development they adopted the CutMix data augmentation strategy, which
on real-time performance, shared control strategies, and was confirmed to enhance model performance (P < 0.05).
asynchronous detection capabilities. Wang et al. [138] con- Ablation experiments further validated the effectiveness
ducted a randomized cross-session online study to evaluate of this strategy across multiple datasets. Additionally, the
the performance of the deep learning model IFNet in 15 authors made the complete model code publicly available
BCI-naive subjects. Experimental results demonstrated that and reported an inference latency of approximately 5 ms,
IFNet significantly outperformed the traditional method, demonstrating the potential of this method for practical
 193 Page 16 of 49 Nano-Micro Lett. (2026) 18:193

deployment. Notably, the study also revealed key limita- inverse reinforcement learning significantly reduced the fre-
tions of the model: offline analysis showed that when the quency of user interventions after only 3–5 demonstrations
model was transferred from calibration tasks to online feed- (p < 0.001). The progress made in this shared control system,
back tasks, significant performance degradation occurred particularly in trajectory distribution reporting, highlights
in both cross-session and cross-mode scenarios; sensitiv- key areas for future breakthroughs. By precisely quantifying
ity analysis by the original authors further indicated that trajectory errors and establishing clear human-machine deci-
when the decision window was shorter than 1 s, the model sion boundaries, future work could substantially improve the
decoding accuracy dropped sharply, even below the chance reliability and safety of shared control paradigms in prac-
level (Fig. 7a). In another study focusing on training effects, tical applications (Fig. 7c). Similarly, focusing on shared
Bhadra et al. [139] employed a rigorous experimental design control, Deng et al. [141] proposed a self-adaptive shared
in a 5-day EEG-based imagined speech BCI study, reveal- control method based on a brain state evaluation network
ing the mechanisms by which training enhances real-time (BSE-NET) for human-wheelchair cooperation. This system
control capabilities. Results demonstrated that after receiv- evaluated the user’s brain control ability online via quantized
ing continuous feedback training, 15 healthy participants attention-gated kernel reinforcement learning and generated
significantly improved their BCI control accuracy from 55 a confidence score to dynamically adjust the control weight
to 70% (p = 0.018), with online cross-validation accuracy between robot autonomy and human operator. Experimental
significantly surpassing the offline condition (T14 = 8.3, results demonstrated that most subjects achieved a task suc-
p = 8.8 × ­10-7, d = 2.14). Pronounced individual differences cess rate of approximately 90% in dynamic environments,
were also observed, as indicated by a significant positive and a significant correlation was observed between the con-
correlation between learning slopes and average perfor- fidence score and EEG decoding accuracy in online experi-
mance (r = 0.55, p = 0.034), corroborating the prevalence of ments. In a complementary approach, Lu et al. [89] proposed
the "BCI illiteracy" phenomenon in imagined speech tasks. a model predictive control (MPC)-based shared control
At the neural mechanism level, the study identified power method. By explicitly setting safety constraints (such as lat-
enhancements in the frontal theta band and the left temporal eral and yaw angle error bounds) and introducing a penalty
low-gamma band as key biomarkers of learning. Technically, term (weight α = 1) for deviations from the driver’s com-
the system achieved an average feedback delay of 100 ms, mand within its optimization framework, this method delin-
and the code and model weights were made publicly avail- eated a quantifiable human-machine trade-off space at the
able. Notably, although EMG signals showed some contribu- system design level. This approach significantly improved
tion to offline decoding (EMG vs. EEG: T14 = 2.2, p = 0.044, task performance: in obstacle avoidance tests, direct brain-
d = 0.57), online EEG decoding accuracy was significantly controlled driving failed. However, with MPC assistance,
higher (T14 = 2.77, p = 0.014, d = 0.71) (Fig. 7b). some participants achieved a 100% task success rate, while
Regarding shared control paradigms, Batzianoulis et al. others attained success rates between 85% and 95%. Future
[140] developed a BCI-based shared control architecture work could focus on conducting multisession, long-term
that integrates the autonomous obstacle avoidance capa- experiments to systematically investigate the system’s adap-
bility of a robot with users’ implicit neural feedback. In tive capabilities in dynamic environments. This approach
this framework, the robot generated real-time trajectories would facilitate the evaluation of the framework’s perfor-
through a dynamical system, while users could exercise mance sustainability under realistic operational conditions.
a "veto" against trajectories that did not align with their Taking a step further, Zhou et al. [142] created a shared
personal preferences by eliciting error-related potentials control system for a 3D robotic arm that integrates a hybrid
(ErrPs). Batzianoulis et al. employed rigorous statistical test- asynchronous BCI (converging SSVEP and electroocu-
ing to validate this paradigm’s effectiveness: they adopted lography (EOG) signals) and computer vision. The sys-
repeated-measures ANOVA, which revealed a significant tem adopted a 15-command asynchronous interface and a
difference in ErrP decoding accuracy between the time- 3D vector synthesis strategy to enable flexible manipula-
locked and continuous modes (F (1, 12) = 27.1, p < 0.001), tion of the robotic arm in unknown environments. Online
with online continuous decoding maintaining an accuracy of experiments demonstrated that in a free spelling task, the
70 ± 13%. Building upon this, posterior probability-weighted system achieved an average classification accuracy of

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 17 of 49 193

92.4% ± 5.5%, a false positive rate of 1.25% ± 0.71%, and negative decisions varied from 0 to 10 (with a maximum
an ITR of 97.9 ± 12.8 bits ­min−1. In the 3D reach-grasp- possible of 34 decisions).In the clinical application of BCIs,
drink task, the shared control mode significantly reduced the Alcaide-Aguirre et al. [146] developed an asynchronous BCI
number of commands issued (9.40 ± 2.14 vs. 28.90 ± 6.83), based on the P300 ERP to facilitate the administration of the
the BCI-guided phase completion time (0.99 ± 0.35 vs. Peabody Picture Vocabulary Test (PPVT-IV) for cognitive
3.04 ± 1.42 min), and the number of error commands assessment in individuals with cerebral palsy. Results from
(0.35 ± 0.59 vs. 2.25 ± 2.02) compared to direct BCI control. the NASA-TLX demonstrated that participants with CP per-
The average trajectory efficiency was 0.80 ± 0.10. They used ceived the BCI-facilitated PPVT-IV as significantly higher
the NASA Task Load Index (NASA-TLX) to assess user in mental demand, physical demand, and effort (p < 0.05)
cognitive load (scores for most dimensions < 30), though compared to their typically developing counterparts.
mental demand and effort scores were relatively higher. In summary, current research has made significant pro-
Future research could focus on quantifying the cumulative gress in enhancing BCI real-time performance, implement-
risks of false positives and false negatives during extended ing intelligent shared control, and constructing robust asyn-
usage periods (ranging from several hours to multiple days). chronous systems. However, a common, unresolved core
Additionally, providing complete distribution data for trajec- issue persists: the vast majority of systems lack an assess-
tory errors and task completion rates would offer valuable ment of their stability and reliability under conditions of
insights for validating system stability and broadening its prolonged (several hours) and multiday continuous use. This
practical applications (Fig. 7d). On the engineering imple- is specifically reflected in the failure to adequately quantify
mentation front of asynchronous BCIs, Hu et al. [143] devel- the cumulative risk of false positives and false negatives,
oped a wearable asynchronous BCI system based on EEG the inability to strictly define the human-machine trade-off
and EOG signals. The system utilized a self-developed com- boundaries in shared control based on complete data dis-
pact amplifier and required only three EEG channels (Cz, tributions (e.g., trajectory error, task completion rate), and
P3, P4) and one EOG channel. It achieved asynchronous the lack of systematic analysis of the long-term impact of
control through the integration of P300 potential detection control parameters (such as penalty weights in model predic-
and blink recognition. In a telephone dialing task, the system tive control). Addressing these issues is crucial for advanc-
attained an average online accuracy of 94.03% ± 4.65%, an ing non-invasive BCIs from the laboratory to real-world
ITR of 31.42 ± 7.39 bits m ­ in−1, and a low false positive rate applications.
of 1.78% ± 2.25% during a 10-min idle state. These promis-
ing results provide a solid foundation for future investiga-
tions into the system’s long-term performance characteris- 2.6 AI‑Enabled Human‑Machine Co‑Adaptive Systems
tics, particularly regarding the cumulative risk profiles of
false positives and false negatives during extended multiday Particularly crucially, current BCI research is undergoing a
usage scenarios. In another asynchronous BCI study, Aloise shift from "open-loop, static" systems to "closed-loop, adap-
et al. [144] designed an asynchronous BCI system based on tive" paradigms [92], with its development direction focus-
ERPs. Online experimental results showed that the correct ing on enhancing real-time decoding capabilities, improving
classification rate of the asynchronous classifier (74.66%) asynchronous detection mechanisms [37], and optimizing
was slightly lower than that of the synchronous classifier shared control strategies [147]. Extending the principle of
(87.96%), and the error rate (7.11%) was also lower than that co-adaptation to resource-constrained settings, Liu et al.
of the synchronous classifier (12.04%). However, neither dif- [148] developed a neuromorphic decoding SSVEP of BCIs
ference was statistically significant (error rate: p = 0.19). The based on a 128,000-cell memristor chip. Their approach fea-
asynchronous classifier demonstrated good robustness dur- tured an interactive update framework that enabled co-evolu-
ing the no-control state, with an average false positive rate of tion between the memristor decoder and the user’s dynamic
0.16 per minute (i.e., fewer than one false positive every six brain signals. This architecture condensed the traditional
minutes). The asynchronous BCI based on SSVEPs designed three-stage pipeline of preprocessing, feature extraction,
by Gernot R. [145]. The participants took 75.5 to 217.5 s and classification into a hardware-friendly one-step opera-
to complete the movement sequence. The number of false tion, reducing computational complexity by approximately
 193 Page 18 of 49 Nano-Micro Lett. (2026) 18:193

Fig. 8  Schematic illustration of AI-enabled human-machine co-adaptive systems. a Left: Schematic diagram depicting the human-machine joint
learning process and the experimental paradigm. Right: Illustration of BCI training with different strategies. Reproduced with permission [148].
Copyright 2025, Springer Nature. b Left: AI copilots use task information to improve BCI performance. Right: CNN-KF architecture. Repro-
duced with permission [149]. Copyright 2025, Springer Nature. c Left: Experimental paradigm. Middle: Effects of online training on MI and
ME robotic finger control. Right: ME and MI tasks using EEG signals filtered with different bandpass settings. Reproduced with permission
[150]. Copyright 2025, Springer Nature

6.5 times while maintaining software-equivalent accuracy. autonomy-based BCI system that enhanced performance
In a brain-controlled drone task with four degrees of free- by introducing an AI copilot. This framework used a CNN
dom, the system demonstrated significant improvements in to extract complex nonlinear features, which were then
energy efficiency and processing speed: decoding consumed passed into a Kalman filter (KF) to predict and correct the
1,643,000 times less energy and achieved a 216-fold increase user’s movement intent. Additionally, the system integrated
in normalized throughput compared to a CPU-based sys- closed-loop decoder adaptation (CLDA), enabling dynamic
tem. A closed-loop co-evolution framework driven by ErrPs optimization of decoding parameters. Experimental results
enhanced decoding accuracy by approximately 20% during a demonstrated that the cursor control system significantly
six-hour online experiment involving 10 participants, estab- improved task efficiency. The average target acquisition rate
lishing a foundational benchmark for low-power, long-term increased by 2.1-fold in healthy participants and 3.9-fold in
adaptive BCI operation (Fig. 8a). a paralyzed participant, with optimized movement trajecto-
Building on this foundation, Lee et al. [149] proposed a ries. In the robotic arm sequential pick-and-place task, the
hybrid adaptive decoding approach and developed a shared paralyzed participant could not complete the task without

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 19 of 49 193

the AI copilot but achieved a 93% success rate with its assis- similarity score of 0.8116 (q < 0.05) across diverse input
tance. Future research could focus on the following promis- modalities. Despite its high performance, decoding accu-
ing directions: developing adaptive decoding algorithms that racy significantly deteriorated when participants engaged
accommodate individual variability and EEG non-stationar- in distraction tasks. These empirical findings highlight the
ity, exploring autonomous perception and decision-making decoder’s sensitivity to user attention and cognitive state,
mechanisms for environments with unknown targets or underscoring the critical importance of behavioral context
dynamic conditions, and validating the system’s generaliza- in human-machine co-adaptive systems.
tion capabilities in more complex application settings. These In summary, AI is driving BCIs toward human-machine
research avenues would facilitate the transition of BCIs from co-adaptive systems capable of dynamic, personalized
controlled laboratory environments to practical real-world interaction [6, 151]. From hardware efficiency to semantic
applications (Fig. 8b). Ding et al. [150] proposed a deep decoding, recent advances demonstrate appreciable poten-
learning model based on the EEGNet architecture, combined tial. Yet, challenges in individual variability, environmental
with a fine-tuning mechanism, facilitating non-invasive real- robustness, task dependency, and cognitive state awareness
time decoding of individual finger MI and motor execution continue to hinder scalability and real-world applicability.
(ME) for robotic hand control. Among 21 able-bodied sub- Future progress will depend not only on better models, but
jects with BCI experience, the system achieved notable on building intelligent ecosystems that are context-aware,
results. Online decoding accuracies reached 80.56% for continuously learning, and capable of seamless cross-plat-
two-finger MI tasks and 60.61% for three-finger MI tasks, form integration.
while the corresponding ME tasks achieved 81.10% and
60.11%, respectively, indicating highly comparable control
3 Integration of Flexible Bioelectronics
performance between the two paradigms. Ding et al. con-
in Non‑Invasive BCIs
firmed through cross-session analysis that the model’s online
accuracy significantly improved with training (two-finger
In contrast to invasive BCIs, which insert probes directly into
MI: F = 7.127, p < 0.001), and applied Bonferroni correc-
target tissue, non-invasive BCIs monitor macroscopic electric
tion in subsequent post hoc analyses to confirm the statisti-
fields outside the cranium to acquire EEG signals. Although
cal significance. The study also revealed clear performance
this approach sacrifices spatial resolution and signal quality,
boundaries: four-finger classification accuracy remained
it avoids the tissue damage caused by invasive techniques
only approximately 46%, insufficient for practical applica-
and potential long-term neuroinflammation. Furthermore,
tion, while the index and middle fingers exhibited the low-
it facilitates rapid and widespread application without the
est decoding performance due to highly overlapping neural
need for complex surgical implantation. These advantages
representations. In the model interpretability analysis, they
have secured long-standing and broad interest in non-inva-
demonstrated through group-averaged saliency topological
sive BCIs [12, 14]. However, their wearable nature presents
maps that the brain regions prioritized by the models exhibit
numerous challenges for EEG acquisition. Firstly, the contact
spatial consistency across different subjects. This indicated
impedance between the electrode and the biointerface signifi-
that the decision-making patterns were grounded in neural
cantly impacts signal quality. Reducing this impedance can
activity features shared across the population. The estab-
be approached from two angles: enhancing the electrode’s
lishment of an adaptive validation framework for complex
intrinsic conductivity and ensuring conformal contact at the
dynamic environments represents a critical breakthrough for
biointerface. Consequently, compared to traditional rigid
future research (Fig. 8c). To expand control granularity, in
electrodes, flexible electrodes not only improve signal quality
the realm of semantic neural decoding, Tang et al. [37] pro-
but also offer greater wearing comfort, facilitating long-term
posed a fMRI-based language reconstruction framework that
monitoring. Flexible electrodes can be fabricated from intrin-
integrates a generative pre-trained transformer model with
sically conductive thin-film materials, such as certain organic
beam search to decode continuous natural language. The sys-
semiconductors, or by integrating conductive nanomaterials
tem bridges internally generated cognition, such as mental
with a flexible substrate [152]. Particularly, driven by inno-
imagery, with stimulus-driven conditions, including speech
vations in materials science, nanomaterials are receiving
or video perception, achieving a peak contextual semantic
 193 Page 20 of 49 Nano-Micro Lett. (2026) 18:193

growing interest and application in flexible electrodes. This with reduced contact impedance and improved SNR in non-
synergy between the distinctive electrical properties of nano- invasive BCIs. Furthermore, mechanical stability plays a key
materials and the favorable mechanical flexibility of sub- role in non-invasive BCIs. Superior mechanical properties
strates provides a robust foundation for high-performance contribute to achieving conformal contact with the target
flexible electrodes [151, 153, 154]. surface, promoting wearing comfort, and maintaining per-
For instance, nanowires (NWs), as one-dimensional mate- formance stability during long-term monitoring while reduc-
rials, possess a high aspect ratio. This structure endows ing the risk of degradation [152]. Due to their nanoscale
them with excellent mechanical flexibility, allowing them dimensions, nanomaterials exhibit excellent robustness.
to maintain structural stability when conforming to irregular They could maintain their performance under repeated
surfaces or under bending conditions [153, 155]. Concur- strain, unlike bulk materials that suffer from performance
rently, due to their efficient electrical transport capabilities, decay due to mechanical fatigue. It is noteworthy that artifact
they serve as conductive fillers that significantly enhance contamination in non-invasive BCI recordings, especially
the electrode’s conductivity [156]. This has been associated during target movement, complicates EEG interpretation.

Fig. 9  Schematic illustration of the performance characteristics of flexible conductive films. a Schematic diagram of the preparation process
for aG-O films. Reproduced with permission [64]. Copyright 2012, American Chemical Society. b Detailed construction of the flexible PENG
device. Reproduced with permission [159]. Copyright 2025, Springer Nature. c Left: A plot of luminance for a foldable ACEL device. Right:
A plot of the luminance stability of it undergoing a cyclic folding test. Reproduced with permission [160]. Copyright 2024, Springer Nature.
d Schematic overview of the gustatory interface. Reproduced with permission [161]. Copyright 2024, Springer Nature. e Left: Flexible sin-
gle-crystalline SRO membranes. Right: Stress relaxation of the film on a water droplet. Reproduced with permission [162]. Copyright 2022,
Springer Nature. f Bending durability of flexible perovskite films. Reproduced with permission [163]. Copyright 2025, Springer Nature

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 21 of 49 193

These artifacts are generally attributable to two origins: bio- another significant advance in flexible optoelectronics. Ma
logical signals and non-biological signals. Non-biological et al. [160] developed 4.5 nm UTAFs, which exhibit 82%
artifacts are largely motion-induced, resulting from electrode average visible transmittance, less than 60 ppm haze, and a
shift that compromises stable interfacial contact. This insta- sheet resistance of 7.6 Ω s­ q−1. These films exhibited excel-
bility manifests as variable contact impedance and intro- lent mechanical stability. The resistance increases by only
duces artifacts. The integration of nanomaterials into flexible 5% under static bending with a curvature radius of 3.5 mm,
substrates presents a promising strategy for ameliorating this and by less than 7% after 100,000 dynamic folding cycles,
challenge. In summary, innovations in materials science are outperforming commercial indium tin oxide electrodes.
essential for achieving higher performance in non-invasive When applied in foldable alternating current electrolumines-
BCIs, and nanomaterials represent a promising candidate cent devices, the UTAFs served as transparent bottom elec-
for this pursuit. trodes to maintain stable luminance, further confirming their
mechanical and electrical stability. These characteristics
suggest that UTAFs may support trends toward improved
3.1 Flexible Conductive Thin Films: Materials
performance when implemented as transparent electrodes in
Innovation and Interface Optimization
flexible optoelectronic devices and intelligent hardware sys-
tems (Fig. 9c). Wang et al. [161] designed a taste interface
The development of non-invasive BCIs is closely linked to
that uses high-density, conformal tongue electrodes to cap-
advancements in flexible conductive materials that strike a
ture electrical signals from the tongue. Combined with BCIs,
balance between electrical properties, mechanical durabil-
ity, and biocompatibility. Recent advances in materials sci- the system enabled taste decoding of a reconstructed tongue
without taste buds, achieving an accuracy of 97.8%. This
ence indicate that conductive films offer viable pathways for
approach provides a novel method for the clinical evalua-
overcoming persistent challenges in signal fidelity, motion
tion and treatment of patients with tongue cancer (Fig. 9d).
artifact suppression, and long-term wearability [157, 158].
Additionally, Kim et al. [162] presented a support-free
For instance, Zhang et al. [64] developed activated reduced
transfer method that facilitates the deposition of ultrathin
graphene oxide (aG-O) films exhibiting in-plane conductiv-
single-crystalline ­SrRuO3 membranes onto flexible poly-
ity of 5,880 S ­m−1. These films exhibit exceptional charge
ethylene terephthalate (PET) substrates, maintaining
transport properties, which are crucial for preserving signal
exceptional structural integrity, optical transparency, and
fidelity in BCI applications (Fig. 9a). Piezoelectric films
electronic performance. These membranes, measuring
represent a valuable class of materials for BCI applications,
approximately 2.5 by 2.5 square millimeters in clean sur-
offering the potential to harvest mechanical energy and
face area and with thicknesses as low as 15 nm, exhibited
potentially power self-sustaining BCI systems. Ren et al.
a high optical transmittance of about 60% in the visible
[159] reported significant progress by developing free-
spectrum. Their electrical resistivity at room temperature
standing (111)-oriented P­ bZr0.52Ti0.48O3 single-crystal thin
ranged between ­10–4 and ­10–3 Ω cm accompanied by robust
films, which combine exceptional mechanical flexibility with
ferromagnetic ordering that persists below 150 K. These
notable piezoelectric performance. These films achieved
SrRuO₃ membranes combine flexibility with functional
an effective piezoelectric coefficient of approximately
performance, suggesting their possible value in develop-
585 pm ­V−1, nearly six times higher than the coefficient
ing flexible electronic and spintronic systems (Fig. 9e).
in their clamped state. The piezoelectric nanogenerators
Jin et al. [163] introduced a spontaneous bifacial cap-
(PENGs) constructed from this material demonstrated an
ping strategy by incorporating 4-(methoxy)benzylamine
ultrahigh power density of about 63.5 mW c­ m−3, maintain-
hydrobromide (MeOBABr) into the perovskite precursor,
ing excellent mechanical resilience with strains exceeding
markedly enhancing the mechanical stability and charge
3.4% and stable output performance after 60,000 continuous
carrier transport properties of flexible perovskite solar
bending cycles. This work not only sets a new benchmark
cells. The work demonstrated that the nanoscale bifacial
for flexible PENG performance but also offers a promising
capping layers formed by MeOBABr effectively planar-
self-powered solution for non-invasive BCIs (Fig. 9b). The
ized grain boundary trenches, mitigated bending-induced
development of ultrathin silver films (UTAFs) represents
 193 Page 22 of 49 Nano-Micro Lett. (2026) 18:193

stress, and improved charge extraction efficiency through In summary, these materials demonstrate the diversity of
surface defect passivation and band alignment optimiza- approaches in developing flexible bioelectronics for non-
tion. The encapsulated target devices retained over 80% of invasive BCIs. Although they use different methodologies,
their initial power conversion efficiency following 10,000 they may exhibit favorable characteristics, including high
bending cycles at a curvature radius of 3 mm. Although conductivity and mechanical compliance that align with
primarily focused on photovoltaic applications, these find- the requirements of wearable BCI design. Future integra-
ings provide valuable insights for material design in flex- tion of these materials into BCI systems could contribute
ible electronics, such as long-term wearable BCIs (Fig. 9f). to enhanced performance of non-invasive neural interfaces.

Fig. 10  Schematic illustration of one-dimensional nanowire structures and properties. a Stability diagram of Ag@Pt AHNWs. Reproduced with
permission [166]. Copyright 2018, American Chemical Society. b Left: Schematic of the structure of stretchable TeNW-TFTs before and after
deformation. Right: Image of TeNW-TFTs under bending. Reproduced with permission [73]. Copyright 2024, The American Association for
the Advancement of Science. c A cyclic bending test result of the flexible RCC electrode at a 15 mm bending radius. Reproduced with permis-
sion [167]. Copyright 2018, John Wiley and Sons. d Left: Band structure of Te illustrating its anisotropic electronic properties. Right: Side-view
SEM image of vertically aligned TeNWs deposited on PI substrate. Reproduced with permission [168]. Copyright 2022, Springer Nature. e A
Schematic of the fabricated fiber-based PD. Reproduced with permission [169]. Copyright 2025, Springer Nature. f Left: Schematic diagram of
the device configuration of the representative self-powered PEC photodetectors based on PbSeNWs. Right: Normalized absorption spectra of
PbSeNWs with different diameters. Reproduced with permission [170]. Copyright 2025, John Wiley and Sons

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 23 of 49 193

3.2 One‑Dimensional Nanowire Bioelectronic Systems increase under compressive testing. Furthermore, under
for Signal Acquisition harsh conditions of 80 °C and 80% relative humidity, the
PUA-coated electrodes maintained stable operation for up
Building on the flexibility and performance characteristics to 240 h, despite a twofold increase in resistance. This level
of conductive films, another emerging material strategy of stability helps meet the need for reliable electrodes in
gaining attention in non-invasive BCI research involves applications that require long-term mechanical flexibility
the utilization of NWs [164]. NWs possess a unique one- and environmental durability (Fig. 10c).
dimensional (1D) topology, high aspect ratio, and tunable In addition to mechanical stability, NWs provide signifi-
surface chemistry. These characteristics demonstrate poten- cant advantages in multimodal integration and functional
tial for achieving the mechanical flexibility, electrochemi- expandability. Li et al. [168] leveraged the intrinsic anisot-
cal stability, and multimodal sensing capabilities required ropy of tellurium NWs to create a bimodal tactile sensor
by BCIs [165]. Moreover, the structural features of NWs (BTS) that achieves the decoupling of pressure and tem-
are believed to facilitate neural signal acquisition: their 1D perature difference signals. The material exhibited a carrier
morphology could promote directional charge transport, mobility of 1,000 ­cm2 ­V−1 ­s−1. When the mechanical pres-
potentially reducing signal attenuation at heterogeneous sure ranged from 0 to 5 kPa mechanical pressure, the sensor
biological interfaces. current increased from an initial 2.82–141.82 μA. Experi-
To overcome the instability of traditional silver nanow- mental results demonstrated that the BTS-based smart glove
ire electrodes, Zhou et al. [166] developed silver-platinum facilitated somatosensory feedback interaction between VR
alloy-walled hollow nanowires (Ag@Pt AHNWs) through and the real world. By integrating sensor signals with deep
electrochemical and galvanic replacement processes. The learning techniques, successful stimulation recognition
silver-platinum alloy enhanced corrosion resistance, while and neural reflex modeling of the rabbit sciatic nerve were
the electrode maintained high optical transparency of 82% achieved. Furthermore, the sensor exhibited excellent bio-
at a wavelength of 550 nm and a low sheet resistance of compatibility, offering broad application prospects in the
28.73 Ω ­sq−1. The alloy structure exhibited excellent ther- biomedical field (Fig. 10d). Karagiorgis et al. [169] engi-
mal stability, sustaining operation at 400 °C for 11 h, and neered a fully transparent and flexible photodetector by inte-
demonstrated robust electrochemical stability in acidic grating ZnO NWs with electrospun PEDOT/PSS/Ag NW-
environments, enabling integration into functional devices based nanofibers on a biodegradable cellulose acetate (CA)
that require long-term reliability. This work provides a scal- substrate. The device demonstrated an ultraviolet responsiv-
able approach for preparing highly stable metal nanomate- ity of 1.10 × ­106 A W
­ −1 under a 5 V bias and 0.5 μW ­cm−2
rial electrodes, which is crucial for flexible optoelectronic ultraviolet (UV) illumination intensity, alongside an optical
devices operating in harsh environments (Fig. 10a). Zhao transmittance of 70% at 550 nm wavelength. This combi-
et al. [73] utilized a biomimetic lock-and-shear assembly nation of high performance and transparency was achieved
strategy to fabricate wafer-scale aligned arrays of tellurium while maintaining robust stability under dynamic UV
nanowires (TeNWs). The thin-film transistors exhibited exposure on flat and curved surfaces. Importantly, the CA
hole mobility exceeding 100 c­ m2 ­V−1 ­s−1 and an on/off substrate and PEDOT/PSS/Ag composite layers exhibited
current ratio approaching 1­ 04 on flexible PET substrates. biodegradability within several months in buffer solutions,
Furthermore, devices fabricated on elastomeric substrates offering an eco-friendly solution to address electronic waste
maintained stable electrical performance under uniaxial challenges. These results underscore the potential of this
strains up to 40%, demonstrating excellent stretchability nanowire-based photodetector platform for next-generation
and mechanical robustness (Fig. 10b). Similarly, Kim et al. wearable and transparent electronic applications (Fig. 10e).
[167] demonstrated that copper nanowire electrodes encap- Li et al. [170] developed sub-1-nm PbSeNWs via a cation-
sulated with polyurethane acrylate (PUA) exhibit significant exchange strategy in N, N-dimethylformamide (DMF) sol-
environmental adaptability and mechanical stability. Spe- vent, which exhibited a near-infrared (NIR) absorption peak
cifically, after 1,500 bending cycles at a radius of 15 mm, centered at 940 nm. These self-powered photoelectrochemi-
the electrodes showed only approximately 3.5% increase cal photodetectors exhibited a responsivity of 113 mA W ­ −1
in resistance under tensile testing and approximately 2.5% and a detectivity of 4.65 × ­1011 Jones without external bias.
 193 Page 24 of 49 Nano-Micro Lett. (2026) 18:193

Fig. 11  Schematic illustration of one-dimensional SiNW structures and properties. a Top: Fin-gate structure with aligned SiNW channels. Bot-
tom: Ids-Vds output and transfer characteristics of SiNW TFTs. Reproduced with permission [175]. Copyright 2023, John Wiley and Sons.
b Left: Pulling stretchable crystalline silicon (c-Si) nanowire springs out of running catalyst droplets. Right: SEM images of SiNWs. Repro-
duced with permission [176]. Copyright 2017, American Chemical Society. c Left: Fabrication of single-nanowire-morphed robotic grippers.
Right: Elastic deformation, structural stability, and force gauging behavior of the nanowire gripper. Reproduced with permission [177]. Copy-
right 2023, Springer Nature. d Top: SEM images of SiNix nanospring arrays. Bottom: Microscopic images showing the geometric evolution of
a SiNix-NS interconnection. Reproduced with permission [178]. Copyright 2021, John Wiley and Sons. e Left: Step-necking growth of ultrathin
and short SiNW channels. Right: SEM image of SiNWs. Reproduced with permission [179]. Copyright 2025, Springer Nature. f Left: SEM
images of SiNHs with forward- or reverse-line feeding growth symmetry. Right: Two distinct resonant frequencies corresponding to the swaging
resonant modes of the SiNHs. Reproduced with permission [180]. Copyright 2020, American Chemical Society

Embedded in flexible polyvinylidene fluoride (PVDF) com- of nanomaterials that effectively combine electrical perfor-
posite films, these NWs combined superior carrier transport mance, mechanical flexibility, and biocompatibility. Silicon
properties with mechanical flexibility. The NIR absorption nanowires (SiNWs) exhibit compatibility with semicon-
at 940 nm aligned with the safety window for human tissue, ductor processing, low-temperature fabrication capabili-
providing a foundation for potential biomedical applications ties, and controllable sub-100 nm architectures, suggest-
(Fig. 10f). ing their potential utility in neural interface applications
Material innovation has contributed to recent progress in [171–173]. These characteristics make SiNWs a potential
non-invasive BCIs, particularly through the development platform for high-density flexible electrode arrays. Studies

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 25 of 49 193

have highlighted their scalable fabrication and favorable or arbitrary two-dimensional patterns by directing catalyst
electrical properties, which may support stable interfac- droplets along predefined atomic step edges. In situ scanning
ing with biological tissues over extended periods [174]. electron microscopy (SEM), tensile testing, and electrical
For example, Song et al. [175] fabricated ultrathin SiNW transport measurements revealed that the c-SiNWs main-
arrays with uniform diameters of approximately 52 nm using tained stable electrical connectivity under tensile strains
an in-plane solid-liquid-solid (IPSLS) growth technique exceeding 200%. This strategy represents a step toward the
conducted below 350 °C. The SiNW thin-film transistors potential development of large-scale, wearable, and bio-
(TFTs) demonstrated a high transparency up to 90% and compliant electronic systems in future (Fig. 11b). Yan et al.
achieved an on/off current ratio greater than 1­ 06. Notably, [177] proposed an ultracompact Ω-shaped robotic gripper
solution-based room-temperature passivation enhanced based on the deformation of a single nanowire and Lorentz
device stability. These advancements provide a foundation force actuation. This gripper was capable of large-amplitude
for high-fidelity neural signal recording in wearable elec- vibration and multidimensional operations, including grasp-
tronics. Nevertheless, to facilitate widespread application ing, lifting, and twisting. Additionally, it achieved precise
in high-density neural interfaces, further improvements in payload release by overcoming van der Waals forces through
the mechanical adaptability and environmental robustness of high-frequency vibration. Paired grippers could collabora-
SiNW arrays are necessary (Fig. 11a). To address this, Xue tively perform complex tasks such as microsphere transfer,
et al. [176] demonstrated a method for deterministic struc- demonstrating strong potential in high-sensitivity biosensing
tural programming of ultralong in-plane SiNWs. Utilizing actuation. Looking ahead, such grippers could be explored
the indium-droplet-guided IPSLS growth mechanism, they as actuators in controlled neuromodulation platforms that
achieved the formation of highly stretchable c-Si springs interface with microscopic biological targets, where high

Table 1  Comparative Analysis of Key Parameters and Impacts of Nanomaterials in Non-Invasive BCIs
Materials Conductivity Young’s Modulus Stretchability Stability Direct Link to Non- References
Invasive BCI Perfor-
mance

AgNWs 2.8 Ω ­sq−1 / Up to 40% ∆Sheet Resistance Excellent signal qual- [181]
≈ ± 50% @ 100 ity; Middle mechani-
cycles for 40% cal durability; Middle
monitoring density
AgNWs 3.2–3.6 Ω s­ q−1 / / 0% increase in frac- Excellent signal qual- [182]
tional resistance dur- ity; Middle mechani-
ing fatigue testing up cal durability; Middle
to 500,000 cycles monitoring density
Nickel silicide NWs 2 × ­106 S ­m−1 170 GPa Up to 50% ∆R/R0 <  ± 1.7% @ Good signal quality; [178]
10, 000 cycles for Middle mechani-
15% cal durability; High
monitoring density
SiNWs Several uA under 170 GPa Up to 45% stable ΔR/R0 @10, 000 Good signal quality; [152]
a bias voltage cycles for 10% Middle mechani-
of ± 3 V cal durability; High
monitoring density
PEDOT/PSS hydrogels 23.7 S ­m−1 8–374 kPa Up to 100% / Good signal quality; [183]
Good mechanical
durability; Middle
monitoring density
d-Sorbitol-PEDOT/ > 1,000 S ­m−1 1.87 MPa Up to 60% Strain-insensitive Good signal quality; [184]
PSS hydrogels resistance after the Good mechanical
initial stretch durability; Middle
monitoring density
 193 Page 26 of 49 Nano-Micro Lett. (2026) 18:193

Fig. 12  The application of zero-dimensional materials in BCIs. a Left: Resistivity of sensors with different graphene content at 2%, 4%, 6%,
8%, and 10%. Middle: Response tests of the film at different working voltages. Right: UV protection mechanism of the coating. Reproduced with
permission [185]. Copyright 2021, John Wiley and Sons. b Left: The SEM images of printed AgNP and AgNP-Ga-In traces. Right: A circuit
on a toy lemon and functioning as an electronic tattoo with an LED on a fingerprint. Reproduced with permission [186]. Copyright 2018, John
Wiley and Sons. c Schematic of the Ag-Au-PtCore-Shell-Shell NWs, TEM images of the control nanocomposite, schematic illustration of the
implantable electrode mounted on the rat’s heart, and comparison of electrogram quality and magnified view of a single cardiac electrogram
peak between Ag NW nanocomposite-based electrode and Ag-Au-Pt NW/Pt NP nanocomposite-based electrode. Reproduced with permission
[187]. Copyright 2023, American Chemical Society. d Schematic of the nanoparticle synthesis process, quantification of the fluorescence inten-
sity of RBITC, and representative confocal fluorescence images of brain slices of rats. Reproduced with permission [188]. Copyright 2022, Else-
vier. e Motion artifacts comparison between the P-P-PDA NP electrode and the standard wet Ag/AgCl electrode, and characterization of PDA
NPs. Reproduced with permission [189]. Copyright 2023, John Wiley and Sons

precision and long-term biocompatibility would be key con- interconnect strategies. After nickel alloying, the electrical
siderations (Fig. 11c). conductivity of the SiNixNWs increased by four orders of
To overcome limitations, Yuan et al. [178] fabricated an magnitude. When integrated onto flexible polydimethylsi-
ordered array of ultrathin, highly conductive Si-Ni alloy loxane (PDMS) substrates, these nanowire springs exhib-
nanowire springs (SiNix-NS) with an average diameter ited excellent mechanical compliance, sustaining tensile
of approximately 160 nm. This innovation allows precise strains exceeding 50% and maintaining stable electrical
spatial control of SiNWs and the programmable design performance over 10,000 cycles at 15% strain. Owing to
of elastic interconnect geometries. This approach is com- their ultracompact architecture and adaptable soft-inter-
patible with existing silicon-based thin-film technologies face compatibility, these nanowire springs demonstrate
and addresses key limitations of conventional stretchable strong potential for low-damage, conformable integration

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 27 of 49 193

in non-invasive BCIs (Fig. 11d). Concurrently, Wu et al. robust environmental adaptability. Notably, the film
[179] used a step-guided necking growth method to fabri- retained its conductivity and self-healing capability even
cate ultrathin, short-necked SiNWs channels, which were after repeated mechanical damage. These attributes suggest
designed for high-performance field-effect transistors potential applications in durable wearable human-machine
(FETs). By tuning the size of indium (In) droplets and the interfaces (Fig. 12a). Furthermore, metal nanoparticles serve
height of the jumping steps, a thick-thin-thick short-chan- as excellent conductive fillers. Their integration with flex-
nel structure could be formed. The necked region could ible substrates combines high electrical conductivity with
be narrowed down to below 25 nm in diameter. These mechanical flexibility. Such materials have been explored
FETs exhibited a steep subthreshold swing of less than as electrode interfaces for non-invasive BCIs, where low
70 mV ­dec−1 and an on/off current ratio exceeding 1­ 07, impedance and stable skin contact are essential for reliable
significantly outperforming counterpart devices with uni- electrophysiological signal acquisition. Tavakoli et al. [186]
form-diameter SiNW channels (Fig. 11e). Ma et al. [180] introduced a room-temperature sintering technique using
demonstrated a novel approach for the three-dimensional eutectic gallium-indium (EGaIn) to significantly enhance
growth and integration of silicon nanohelices (SiNHs) on the electrical and mechanical properties of inkjet-printed
the sidewalls of bamboo-like cylindrical structures. In silver nanoparticle (AgNP) traces for stretchable thin-film
this method, periodic sidewall grooves were first formed electronics. In this method, AgNPs with diameters around
via Bosch etching, and indium catalyst droplets were then 100 nm were printed onto temporary tattoo paper and sub-
employed to guide the helical growth of silicon along sequently coated with a thin layer of liquid-phase EGaIn.
these grooves. By tuning growth parameters, the diameter, This process induced the aggregation of AgNPs and filled
pitch, aspect ratio, and chiral/achiral symmetries of SiNHs microcracks, forming a continuous conductive Ag-Ga-In
could be precisely controlled. Furthermore, the SiNHs can composite trace without requiring high-temperature treat-
be reliably released as individual units (Table 1). These ment. The incorporation of EGaIn increased the electrical
SiNHs exhibit structural programmability and multifunc- conductivity of the printed traces by six orders of magni-
tionality that may be advantageous for flexible bioelec- tude, reaching 4.85 × ­106 S m
­ −1, and improved the maximum
tronic systems, including non-invasive BCIs (Fig. 11f). failure strain from 4.5% to over 118%. The AgNPs played
a critical role in forming conductive percolation networks
3.3 Zero‑Dimensional Materials in BCIs: Innovation and, upon interaction with EGaIn, contributed to the for-
and Application mation of heterogeneous Ag-In-Ga clusters that enhanced
both conductivity and mechanical stretchability. The result-
Zero-dimensional materials, such as nanocapsules and ing circuits exhibited stable performance under strains up
nanoparticles, often exhibit size-dependent properties dif- to 80%, low electromechanical coupling (gauge factor ≈1),
ferent from those of their bulk counterparts, owing to their and compatibility with hydrographic transfer to complex 3D
nanoscale dimensions. These distinctive characteristics are surfaces and human skin. This approach facilitates the fab-
of significant importance for non-invasive BCIs. For non- rication of highly deformable and robust electronic tattoos
invasive BCIs intended for long-term monitoring, the deg- and stretchable sensors, offering a practical route toward
radation of electrode materials significantly impacts signal skin-conformal and shape-adaptive electronics (Fig. 12b).
acquisition quality. UV radiation is a critical factor in the Similarly, Sunwoo et al. [187] designed a stretchable low-
aging process, particularly for certain organic flexible films. impedance conductor based on Ag-Au-Pt core-shell-shell
To address this issue, Zhou et al. [185] developed a multi- NWs and in situ synthesized platinum nanoparticles (Pt
functional self-healing hybrid film composed of titanium NPs) embedded in a styrene-ethylene-butadiene-styrene
dioxide nanocapsules, graphene, and multibranched polyu- (SEBS) elastomer. The Ag-Au-Pt NWs feature a conduc-
rethane, exhibiting excellent mechanical flexibility and UV tive Ag core, a biocompatible Au inner shell, and a low-
protection properties. Within an optimized graphene content impedance Pt outer shell with a highly embossed structure
range, the film maintained stable electrical properties, dem- that enhances the effective surface area and charge transfer
onstrated a sensitive piezoresistive response, and possessed efficiency. The in situ formed Pt NPs, uniformly dispersed
 193 Page 28 of 49 Nano-Micro Lett. (2026) 18:193

during composite fabrication, serve as critical conductive recording and high-accuracy sustained attention evaluation.
bridges between the NWs, reinforcing the percolation net- The hydrogel is constructed using biocompatible polyvinyl
work. This synergistic structure yields a nanocomposite with alcohol (PVA) and polyvinylpyrrolidone (PVP) via a ketali-
high conductivity (~ 11,000 S ­cm−1), exceptional stretch- zation reaction, resulting in a soft network with tissue-like
ability (~ 500%), low electrochemical impedance (166.5 Ω modulus, high transparency, and excellent flexibility. To
at 1 kHz), and significantly reduced cytotoxicity due to the enhance functionality, polydopamine nanoparticles (PDA
effective suppression of Ag ion leaching. The incorporation NPs) are incorporated into the hydrogel matrix through
of Pt NPs is essential for simultaneously enhancing electri- an oxidative degradation process, which converts opaque,
cal, mechanical, and electrochemical properties, supporting micron-sized PDA aggregates into transparent, nanosized
their use in high-performance wearable biosignal monitoring particles. The introduction of PDA NPs improves the
systems (Fig. 12c). hydrogel’s self-adhesiveness, conductivity, and interfacial
Beyond their role as conductive fillers, nanoparticles of compatibility while maintaining high optical transparency
certain materials exhibit unique enhancements in electro- and biocompatibility. The resulting multichannel electrode
magnetic fields. For instance, iron oxide nanoparticles dem- exhibits low interfacial impedance, high channel uniformity,
onstrate superparamagnetism, which amplifies the electric low noise power, and robust performance under motion and
field induced by the external field near the neuronal mem- sweating conditions. Moreover, the system demonstrated the
brane, thereby exciting neuronal activity more effectively. ability to classify prefrontal EEG signals into seven levels
Hong et al. [188] developed a non-invasive strategy for of sustained attention with 91.5% accuracy using a linear
enhancing transcranial magnetic stimulation (TMS) using support vector machine (LSVM) classifier. This nanoarchi-
tailored superparamagnetic iron oxide nanoparticles (SPI- tectonics strategy highlights the critical role of PDA NPs in
ONs) to improve functional recovery after ischemic stroke. enabling high-performance, multifunctional hydrogel elec-
They synthesized water-soluble Tat peptide-conjugated trodes for personalized health monitoring and cognitive state
SPIONs (Tat-SPIONs) coated with chitosan and polyeth- assessment (Fig. 12e).
ylene glycol, which exhibit excellent colloidal stability and
superparamagnetism. A combination of intranasal adminis-
tration and external magnetic guidance facilitates the effi- 3.4 Wearable Flexible Devices: Design Principles
cient, non-invasive delivery of these nanoparticles across and Functional Integration
the blood-brain barrier and into the brain parenchyma of
rats. The delivered Tat-SPIONs significantly enhance the 3.4.1 Design Principles
neurostimulatory effects of TMS, as evidenced by increased
motor-evoked potential amplitudes, reduced motor thresh- As the development of non-invasive BCIs progresses, the
olds, elevated c-fos expression, and marked improvements integration of wearable flexible devices has become a key
in motor-sensory and cognitive functions in a stroke model. area of focus. These devices aim to bridge the gap between
Mechanistic investigations revealed that the enhancement advanced bioelectronic systems and user-friendly, comforta-
was primarily mediated by a highly localized magnetoelec- ble applications [154]. To achieve this, the design of flexible
tric effect from the plasma membrane-associated nanopar- electronics must strike a delicate balance between mechani-
ticles, which amplified the TMS-induced electric field to cal flexibility, biocompatibility, and signal fidelity-ensuring
trigger neuronal activation. This platform demonstrates a that devices can comfortably interface with the human body
viable pathway for medical translation of nanomaterial- while maintaining high performance over extended periods
enabled remote brain stimulation (Fig. 12d). What’s more, of use [190, 191]. The rapid evolution of flexible bioelec-
certain nanoparticles, due to their hydrophilic/hydrophobic tronics is driving advances in the landscape of wearable
properties, hold significant potential for application in the BCIs, enabling opportunities for real-time monitoring and
interface engineering of non-invasive BCIs. This can be seamless interaction with dynamic biological systems. Sili-
leveraged to ensure more intimate contact between the elec- con is the most mature and reliable semiconductor material
trode and the skin. Han et al. [189] developed a flexible and in current manufacturing processes, and is widely used in
self-adhesive hydrogel electrode for long-term wireless EEG electronics and sensing. However, integrating silicon with

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 29 of 49 193

Fig. 13  Wearable flexible devices: design principles. a Magnified photo of the hard island region at various strains on the PDMS film and the
array attached to human skin. Reproduced with permission [192]. Copyright 2022, John Wiley and Sons. b Top: Optical image of a flexible
SiNW strain sensor array attached to a human finger. Bottom: Normalized current maps of a sensor matrix under varied strains and its real-time
response. Reproduced with permission [193]. Copyright 2025, American Chemical Society. c Schematic illustration of the pickup and printing
process of PLMT. Reproduced with permission [194]. Copyright 2024, Springer Nature. d Left: 3D MMF with serpentine wires. Right: Simu-
lated 3D distributions of light and temperature on neural spheroids. Reproduced with permission [195]. Copyright 2021, The American Associa-
tion for the Advancement of Science

flexible substrates to extend its applications into flexible 50% tensile strain and supported 1,000 mechanical cycles at
electronics poses a significant challenge. As the fundamen- 20% strain, proving the robustness and durability of these
tal building block of modern integrated circuits, transistors devices. These SiNWs FETs exhibited a hole mobility of
hold great potential in non-invasive BCIs due to their gating 70 ­cm2 ­V−1 ­s−1, an on/off current ratio exceeding ­105, and
and current amplification capabilities. Yet, limited by the subthreshold swing values ranging from 134 to 277 milli-
intrinsic brittleness of silicon, transistors on flexible films volts per decade. Notably, the devices demonstrated stable
often suffer structural damage due to modulus mismatch. To electrical operation over a period of 270 days under ambient
address this, Song et al. [192] directly integrated transistors conditions without any encapsulation or passivation layers.
onto a flexible film using a strategy based on rigid-island Optical and scanning electron microscopy images confirmed
protection. This design maintained functional integrity under conformal skin adhesion during deformation, while finite
 193 Page 30 of 49 Nano-Micro Lett. (2026) 18:193

element analysis revealed that the stress-optimized island In non-invasive BCIs, modulus matching at the biointerface
layout effectively minimized interfacial strain, supporting is crucial. Matched Young’s modulus tends toward achieving
long-term mechanical stability (Fig. 13a). Furthermore, conformal contact with the epidermis, suggesting the potential
integrating nanomaterials with flexible films to combine to reduce contact impedance, improve the SNR, and mitigate
mechanical flexibility and electrical conductivity is a widely motion-induced artifacts. Such mechanical compatibility could
adopted strategy. This is often achieved by directly transfer- also contribute to improved comfort during extended wear. For
ring nanomaterials onto flexible films using a post-transfer instance, designs such as wavy, serpentine, and fractal struc-
technique. However, this method faces challenges such as tures facilitate better modulus matching with the biointerface
the difficulties in precise integration, potential device non- and potentially enhance the mechanical stability of non-invasive
uniformity, and low pixel density. Consequently, the chal- BCIs under dynamic conditions. In scenarios requiring micro-
lenge of fabricating high-density, precisely integrated chips, scale 3D neural monitoring, traditional thin-film electrodes
akin to those produced by top-down etching processes like often fall short. Therefore, structurally designing electrodes
photolithography, presents a significant hurdle for develop- to achieve conformational matching with the target measure-
ing higher-density, higher-performance non-invasive BCIs. ment sites is of great importance for the advancing field of BCI
To address this, Song et al. [193] directly grew elastically technology. Park et al. [195] proposed a three-dimensional
deformable SiNWs on a flexible polyimide film and inte- multifunctional neural interface platform designed for cortical
grated them into high-performance strain sensors. This spheroids and engineered neural assembloids. This platform
architecture combines structural elasticity and conformabil- leverages mechanically guided assembly techniques to trans-
ity, offering a robust platform for continuous biomechanical form planar structures into 3D architectures, enabling multi-
monitoring in soft-interface neuroelectronics (Fig. 13b). functional interactions with neural spheroids. Utilizing ultrathin
Beyond the direct fabrication and integration of electronic polyimide substrates and microscale gold electrodes, the sys-
devices on flexible films, transfer printing also serves as a rapid tem achieved high-fidelity detection of action potentials, with
and effective strategy. However, when transferring exception- an average signal duration of 0.5 ms and peak-to-peak ampli-
ally thin electronic devices or electrodes onto flexible films or tudes reaching 15 microvolts. The platform exhibits favorable
biological interfaces, issues such as structural fractures due to mechanical compliance, with relatively low bending stiffness
stress concentration or delamination caused by poor interfa- that facilitates conformal contact with soft biological tissues.
cial adhesion may arise. To address this challenge, Shi et al. Its elastic interconnects remain functional under small applied
[194] developed a phase-transition-enabled gallium stamp strains, enabling stable and reasonably conformal integration
that facilitates damage-free, three-dimensional transfer print- near the cortical surface. This research also demonstrated the
ing of ultrathin functional membranes. This method achieved potential platform’s multifunctionality across electrical, optical,
seamless integration onto complex curved surfaces, including thermal, and chemical modalities, with experimental validation
fingernails and contact lenses, with an average displacement of its capabilities in monitoring neural activity and investigating
across the array of 21.9 µm. By harnessing gallium’s reversible neural injury and recovery (Fig. 13d).
solid-liquid-phase transition, the stamp applied an exceptionally
gentle preload of 0.0053 MPa while achieving robust adhesion
up to 0.15 MPa. Notably, the approach reduced interfacial shear 3.4.2 Functional Integration
strain by four orders of magnitude compared to conventional
PDMS stamps, significantly mitigating mechanical mismatch Building on these advances in flexible electronics, the next
and delamination risk. These capabilities highlight the technol- step is applying them specifically to non-invasive BCIs.
ogy’s potential for integration into non-invasive EEG systems, With the advancement of non-invasive BCI technology,
where mechanical stretchability and long-term skin conform- EEG monitoring is no longer confined to short-term, static
ance are essential for stable signal acquisition over irregular observations, as long-term dynamic monitoring is garner-
scalp geometries (Fig. 13c). Structural engineering, as one of ing increasing attention. However, when BCIs monitor mov-
the most widely adopted strategies in flexible electronics, also ing subjects, new challenges are posed to the mechanical
holds significant potential for non-invasive BCIs. It can partially durability of electrodes. Prolonged dynamic operation could
compensate for the limitations of materials’ intrinsic properties. introduce structural defects from mechanical fatigue within

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 31 of 49 193

Fig. 14  Wearable flexible devices: functional integration. a Compatibility of TTE with MEG. Reproduced with permission [196]. Copyright
2020, Springer Nature. b Left: Schematic diagram of EEG signal monitoring. Right: EEG signals recorded with corresponding spectrogram in
the eyes-closed state. Reproduced with permission [197]. Copyright 2024, John Wiley and Sons. c Top: Impedance map at 1 kHz for all elec-
trodes on the subject’s head. Bottom: a 3-T clinical scanner. Reproduced with permission [198]. Copyright 2021, The American Association
for the Advancement of Science. d Left: Contact analysis and EEG signal recording. Right: The SEM images of the AgPMS. Reproduced with
permission [199]. Copyright 2019, American Chemical Society. e Top: Design of in-ear bioelectronics. Bottom: Schematic diagram of the cock-
tail effect experiment based on SpiralE. Reproduced with permission [9]. Copyright 2023, Springer Nature. f Left: Experimental setup for EEG
acquisition and drowsiness labeling. Right: Earpiece assembly and fit. Reproduced with permission [28]. Copyright 2024, Springer Nature

the electrodes, leading to a degradation in their EEG sig- relative to conventional electrodes. TTEs also conformed
nal detection capability. Consequently, materials capable of closely to the skin surface and were less affected by hair,
self-recovery hold significant potential for enabling stable, making them well-suited for long-term monitoring. There-
long-term monitoring with non-invasive BCIs. Ferrari et al. fore, this technology holds promise for potential applications
[196] fabricated ultrathin conductive polymer tattoo elec- in multimodal brain monitoring and diagnostics (Fig. 14a).
trodes (TTEs) using inkjet printing technology and validated Simultaneously, relative motion between a non-invasive
their performance in clinical EEG as well as their compat- BCI and the body interface can introduce artifacts during
ibility with MEG. TTEs were able to detect alpha waves measurement, which severely compromise the quality of
with significantly higher signal amplitude around 20 Hz the acquired signals and complicate the extraction of target
compared to Ag/AgCl electrodes, demonstrating greater neural information. Therefore, maintaining stable contact
sensitivity. Moreover, during auditory evoked potential between the non-invasive BCIs and the target interface is
recordings, TTEs exhibited a superior SNR (4.07 vs. 3.36) paramount. Bioadhesive materials, capable of conforming
 193 Page 32 of 49 Nano-Micro Lett. (2026) 18:193

intimately to biological surfaces, offer low interfacial imped- motor-related mu suppression signals, with signal quality
ance and favorable mechanical flexibility. This presents an comparable to clinical-grade gelled electrodes. The devices
effective strategy for achieving high SNR and stable moni- exhibited excellent mechanical compliance, scalable laser-
toring with non-invasive BCIs. Wang et al. [197] designed based fabrication, and strong compatibility with MRI and
an epidermal sensor based on bioadhesive MXene hydrogel, CT imaging, collectively contributing to the mitigation of
in which a dynamic cross-linked network provided excel- several key challenges in wearable neurotechnology. This
lent conductivity retention under 200% tensile strain and work establishes MXene composites as a key materials plat-
rapid mechanical self-healing. These properties contribute form capable of advancing adaptive, multimodal bioelec-
to maintaining stable performance in scenarios involving tronic systems that bridge clinical-grade and consumer-grade
complex skin deformations. By synergistically integrating BCIs (Fig. 14c). Among the numerous parameters influenc-
conductive MXene nanosheets with bioadhesive functional ing BCIs, hair and sweat are notably significant sources of
groups, the hydrogel interface achieved tunable adhesion interference. Hair could impede intimate contact between
strength ranging from 10.17 to 38.75 kPa, maintaining the electrode and the scalp, increasing contact impedance,
stable electrode-skin impedance during motion and sup- reducing the SNR, and amplifying motion-induced artifacts.
porting high-fidelity EEG signal acquisition. The three- Sweat, due to its ionic content, alters the direct interfacial
dimensional porous architecture offered not only ultraviolet impedance, leading to DC drift and slow-wave artifacts. To
shielding but also antibacterial properties, with inhibition mitigate the interference from hair, strategies can be devised
rates of 89.61% and 93.15% against Escherichia coli and from the perspectives of electrode structure and material.
Staphylococcus aureus, respectively, ensuring long-term For instance, replacing planar film electrodes with spike-
wearability. Integrated with a machine learning algorithm, shaped electrodes can effectively bypass hair and establish
the system achieved an accuracy of 98.1% in EMG-based stable contact directly with the scalp. Alternatively, employ-
sign language recognition. This multifunctional platform ing porous materials like a sponge as the electrode substrate
holds great potential for advancing next-generation wearable allows it to compress around hair follicles under application
electronics and machine learning-assisted human-machine pressure, maintaining consistent contact with the scalp. Lin
interaction (Fig. 14b). The electrodes used in non-invasive et al. [199] developed a flexible, gel-free silver nanowire-
BCIs can be simply classified according to material proper- based sponge electrode (AgPMS), which exhibits high con-
ties into three types: wet, semi-dry, and dry. Wet electrodes ductivity and excellent mechanical and chemical stability,
have been extensively applied in clinical practice. A key allowing for effective contact with the skin, bypassing hair.
advantage lies in their use of an electrolyte medium that fills In SSVEP experiments, AgPMS achieved a classification
interfacial gaps on the skin, enabling effective contact. This accuracy of 86% on hairless skin and 82% on hairy skin,
mechanism facilitates low interface impedance and allows a performance closely approaching the 88% accuracy of
for stable signal acquisition with a high SNR. However, the standard gel electrodes. These results demonstrated that
intrinsic properties of wet electrodes render them unsuit- this electrode can significantly improve the performance of
able for long-term monitoring. On the one hand, the con- gel-free, non-invasive BCIs, providing a superior option for
ductive gel applied at the biointerface to facilitate contact EEG recordings in applications such as assistive devices for
can dry out over extended periods, leading to a significant individuals with disabilities and mental state monitoring.
degradation of signal acquisition capability. On the other It holds significant clinical and practical value, particularly
hand, wet electrodes can cause subject discomfort, and pro- for real-world, hair-covered non-invasive BCI applications
longed skin contact with the moisture may induce irritation (Fig. 14d).
or inflammation. Hence, dry electrodes are regarded as well- Emerging ear-centric neural interfaces capitalize on the
suited for long-term non-invasive BCI monitoring applica- ear’s unique anatomical advantages, such as hairless regions
tions. Driscoll et al. [198] developed MXtrodes, a gel-free, for stable electrode contact, inherent mechanical stability
MXene-based bioelectronic platform for high-resolution that helps mitigate motion artifacts, and compact form fac-
electrophysiological recordings in neural and neuromuscular tors compatible with wear. These combined features indi-
systems. In non-invasive EEG applications, 3D pillar-shaped cate long-term neural monitoring. Wang et al. [9] developed
MXtrodes achieved precise acquisition of alpha rhythms and SpiralE, an in-ear bioelectronic platform that conforms to

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 33 of 49 193

the shape of the ear canal. It autonomously expands and (≈6.9 kPa) and high stretchability (≈690%) ensure confor-
spirals along the ear canal under electrothermal actuation, mal contact with the skin during deformation, significantly
ensuring optimal contact with the ear canal for reliable EEG reducing contact impedance compared to commercial Ag/
sensing. This makes it suitable for the construction of both AgCl electrodes and Ecoflex-ECC electrodes, along with
visual and auditory BCIs. In a 9-target SSVEP BCI classifi- minimizing signal drift and noise. Consequently, the AHBH-
cation, offline accuracy reached 95%, and online accuracy in ECC electrode exhibits exceptional electrical stability in
a calibration-free 40-target SSVEP speller experiment was dynamic environments such as vibration and sweating condi-
75%. For auditory attention decoding in cocktail party sce- tions, with its noise RMS value measuring only 9.5 ± 1.6 μV.
narios, natural speech classification accuracy achieved 84% This performance is substantially superior to that of Ag/
(Fig. 14e). Complementing this, Kaveh et al. [28] developed AgCl electrodes (46.3 ± 4.6 μV) and Ecoflex-ECC electrodes
a wireless dry-electrode in-ear EEG monitoring system for (93.0 ± 7.8 μV), indicating enhanced resistance to motion
drowsiness detection. During 35 h of drowsiness monitoring artifacts. The high-quality raw signals facilitate the imple-
across nine participants, the support vector machine model mentation of more complex algorithms: the study introduced
achieved an average drowsiness event detection accuracy a combination of differential entropy and power spectral
of 93.2% for old users and 93.3% for new users (Fig. 14f). density features to enhance emotional state differentiation
These results validate the feasibility of wireless dry user- and employed a domain adaptation neural network to com-
generic earpieces for drowsiness classification and lay the pensate for residual hardware noise and cross-subject vari-
groundwork for population-trained classification in future ability. This approach ultimately increased the average emo-
electrophysiological applications. tion classification accuracy from approximately 85% with
traditional neural networks to 90%. An important direction
for future research is to quantify the marginal improvement
3.5 Integration of Flexible Electronics and Deep in final classification accuracy attributable to the hydrogel
Learning for Advanced Non‑Invasive BCIs interface through direct comparison with conventional elec-
trodes (Fig. 14a).
Recent performance gains in non-invasive BCIs have been Extending the co-design concept to system-level inte-
enabled by the co-design of flexible bioelectronics and deep gration, Mahmood et al. [201] developed a wireless soft
learning-based decoding algorithms. This synergistic para- bioelectronic BCI platform that integrates dry microneedle
digm establishes a positive feedback loop: advanced materi- electrodes, stretchable interconnects, and a VR headset.
als and interface engineering provide a high-fidelity signal The interconnect structure demonstrates negligible resist-
foundation for algorithms, while intelligent algorithms, in ance change after 100 cycles at 50% strain, ensuring signal
turn, compensate for the inherent physical limitations of stability during dynamic movement. At the signal level,
hardware and guide the design direction of future hardware, this architecture effectively suppresses motion artifacts
collectively pushing the boundaries of system performance. and electrostatic interference, thereby enhancing both the
Building upon material innovations, Yang et al. [200] SNR and recognition performance of SSVEP. For fea-
developed a biosensor based on an adhesive-hydrophobic ture extraction and algorithm implementation, the study
bilayer hydrogel (AHBH), which was integrated into a port- introduced a split-eye asynchronous stimulation (SEAS)
able head-mounted device for high-fidelity EEG-based emo- paradigm that expands the number of distinguishable fre-
tion classification. Serving as the core interface, the AHBH quency combinations to 32, subsequently employing a spa-
material leverages a mussel-barnacle-inspired bioadhesion tial convolutional neural network (Spatial-CNN) to achieve
mechanism combined with a hydrophobic surface reorgani- high-resolution classification within a short 0.8 s time
zation strategy. This design achieves a dry-state adhesion window. Under a configuration of 33-class stimuli and
strength of 59.7 N m⁻1 and a water contact angle of 133.87°, 4-channel acquisition, the system achieved a classification
effectively blocking sweat penetration, suppressing motion accuracy of 78.93% ± 2.36% and an ITR of 243.6 ± 12.5
artifacts, and maintaining an adhesion strength of 40.8 N bits ­min−1, significantly outperforming the traditional Ag/
m⁻1 after 20 peeling cycles, thereby demonstrating excellent AgCl electrode system, which achieved an accuracy of
skin adhesion stability. The material’s low elastic modulus 74.72% ± 3.03% and an ITR of 222.4 ± 15.0 bits ­min−1.
 193 Page 34 of 49 Nano-Micro Lett. (2026) 18:193

The system was demonstrated in real-time applications, below 20 kΩ, thereby significantly suppressing signal
including VR text spelling and virtual environment naviga- artifacts and electromagnetic interference caused by rela-
tion. None of the nine subjects reported discomfort during tive interfacial motion. The system enables high-quality,
one hour of continuous use, indicating satisfactory wear- stable EEG signal acquisition, providing reliable input for
ability and user experience. Future studies building on subsequent deep neural network processing. Quantitative
this work could expand participant diversity to include experiments demonstrated that the system achieves an
individuals with motor disabilities and evaluate long-term SNR of 46.6 ± 2.16 dB in SSVEP detection, performing
wearing comfort through extended-use trials (Fig. 14b). significantly better than traditional gel electrode systems
Further advancing the hardware-algorithm closed-loop (16.94 ± 4.60 dB) and existing portable wireless systems
integration, the same research team reported the SKIN- (28.89 ± 2.28 dB). Concurrent reliability tests confirm
TRONICS system [10]. This system utilizes flexible cir- excellent long-term stability: the flexible circuit exhib-
cuits and dry electrodes composed of ultrathin aerosol-jet its resistance changes of less than 0.09 Ω after repeated
printed skin electrodes and flexible conductive polymer, 180° bending around a radius as small as 1.3 mm, while
with a mechanical modulus highly matched to scalp tis- maintaining stable wireless signal strength within a 15-m
sue to achieve conformal contact. This configuration range. Additionally, the elastomer hairy electrode shows
effectively reduces electrode-skin contact impedance to less than 10% change in contact impedance after 1,000

Fig. 15  Integration of flexible electronics and deep learning for advanced non-invasive BCIs. a Schematic illustration and performance of the
AHBH skin-interfaced biosensor for human emotion classification. Reproduced with permission [200]. Copyright 2022, John Wiley and Sons.
b Schematic of VR-enabled portable BCIs. Reproduced with permission [201]. Copyright 2022, Elsevier. c Schematic of the system architecture
featuring fully portable and wireless scalp electronics. Reproduced with permission [10]. Copyright 2019, Springer Nature

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 35 of 49 193

compression cycles. On the algorithmic side, a grid-search characteristics and algorithmic computational requirements
optimized dual-layer CNN enables automatic extraction from the initial design stages, while avoiding latency and
of robust features directly from two-channel time-domain privacy risks associated with cloud transmission. Ultimately,
SSVEP signals, effectively overcoming the performance translating these laboratory innovations into reliable user
limitations of traditional methods with limited channels. assistive tools demands rigorous validation in real-world
The cross-subject generalization capability of the dual- environments. The principal future challenge involves sys-
layer CNN was validated through a six-fold cross-subject tematically evaluating the decoding robustness, long-term
validation, where models trained on 5 subjects were tested wear comfort, and biocompatibility of these co-designed
on an independent subject. Benefiting from this hardware- systems across diverse populations during continuous multi-
software co-optimization, the system achieved an offline day monitoring, while simultaneously addressing accom-
classification accuracy of 94.54% ± 0.9% and an ITR of panying system-level engineering issues to complete the
122.1 ± 3.53 bits ­min−1 using only two EEG channels. The crucial transition from technological prototype to practical
system successfully translates decoded features into con- product.
trol commands, enabling real-time, precise control of an
electric wheelchair, a wireless vehicle, and demonstration
software (Fig. 14c). 3.6 Systems‑Engineering for Non‑Invasive BCIs
In summary, these innovative approaches collectively
outline a clear trajectory for synergistic development in In multisubject collaborative non-invasive BCI systems, pre-
non-invasive BCIs. Advancements in flexible electrodes cise clock management poses a critical challenge for ensur-
establish the hardware foundation by optimizing interfacial ing data timeliness and comparability. The core issue stems
contact and suppressing motion artifacts, thereby consist- from inherent discrepancies in the internal clock sources
ently providing more stable, high-SNR raw signals that sup- of independent acquisition devices. Due to manufacturing
ply high-quality, high-fidelity data for subsequent decoding. tolerances of crystal oscillators and of environmental influ-
Building upon this foundation, advanced algorithms such ences such as temperature and voltage fluctuations, these
as deep learning leverage their powerful feature extraction clocks exhibit varying degrees of drift and offset. Conse-
capabilities to not only achieve significant improvements quently, temporal misalignment gradually develops across
in decoding performance compared to traditional methods devices, making it difficult to reliably synchronize neural
but also actively compensate for inherent hardware limita- responses to the same stimulus from different subjects. This
tions, including residual noise and cross-subject variabil- synchronization problem is especially pronounced in medi-
ity. This synergistic paradigm, embodying the principle of cal applications where communication firmware in clinical-
hardware establishing the foundation and algorithms driving grade devices is often closed and restricted by regulatory
advancement, creates a positive feedback loop that collec- requirements, preventing access to precise timestamps at
tively expands the performance boundaries of non-invasive the hardware level; timestamps can only be recorded at the
BCIs (Fig. 15). application layer, introducing substantial indeterministic and
Looking ahead, advancing beyond current performance variable software and communication delays, which signifi-
limitations requires the development of specialized hard- cantly increase synchronization complexity. A representa-
ware designed for new computing paradigms. The traditional tive collaborative solution combines dedicated hardware
von Neumann architecture encounters energy efficiency modules with synchronization software protocols. A syn-
bottlenecks when processing neural signals, whereas neu- chronization hub is deployed within the network alongside
romorphic computing approaches, such as those utilizing multiple distributed synchronization plugins. Through peri-
memristor-based in-memory computing chips combined odic reference message broadcasting, timestamp exchange
with event-driven spiking neural networks, offer a promis- among nodes, and dynamic reference node election, each
ing solution for ultra-low-power edge computing [151]. This node applies linear regression to continuously estimate and
calls for hardware-software co-design, involving cross-level compensate for its own clock offset relative to the reference
optimization from materials and devices to algorithms. Such node. This allows the system to maintain network-wide
co-design ensures deep integration of hardware physical temporal consistency even under unstable communication
 193 Page 36 of 49 Nano-Micro Lett. (2026) 18:193

latency conditions [202]. When hardware-level alignment is harmonics, as well as noise from consumer electronics such
infeasible, software frameworks, such as the lab streaming as Wi-Fi and Bluetooth devices. Additionally, physiological
layer, could achieve network-based timestamp calibration artifacts generated by the user—such as EOG or EMG—can
[87, 88], while dynamic time warping and event resampling have amplitudes much larger than EEG signals. These dis-
could correct residual drift during post-processing. Future turbances can be introduced into the system through elec-
systems will require synchronization mechanisms that com- trodes or leads acting as "antennas", which deteriorate signal
bine millisecond-level real-time performance, cross-device quality, distort data, and may result in misclassification or
stability, and cross-session reproducibility to support high- system malfunction. This presents direct risks to the reli-
precision applications such as clinical diagnostics, collabo- ability and safety of BCI systems in applications such as
rative BCIs, and large-scale hyperscanning studies. monitoring, diagnosis, or neuro-assistive tasks. To address
In non-invasive BCIs such as EEG, power consumption this complex issue and ensure stable system performance in
and thermal management represent major challenges [203]. real electromagnetic environments, a comprehensive engi-
Although such systems do not require surgical implantation, neering strategy is required. This strategy must integrate
their power dissipation may still lead to localized tempera- hardware design, signal processing, and preliminary evalu-
ture increases, affecting user comfort and long-term safety ation [205]. On the hardware side, shielded cables, driven
[204]. Because EEG electrodes are in direct contact with the right-leg circuits, and other active shielding techniques are
scalp—where blood perfusion is relatively low and thermal used to minimize electromagnetic penetration, while ampli-
conductivity is poor—heat generated by signal acquisition, fiers with high input impedance and high common-mode
amplification, or wireless transmission modules is difficult rejection ratios are employed to enhance the detection of
to dissipate efficiently. This may result in local tempera- weak differential signals. On the algorithmic side, residual
ture elevation that approaches or exceeds biological safety noise persists despite hardware optimization, necessitating
thresholds. Furthermore, non-invasive devices are typically advanced signal processing techniques such as blind source
portable and battery-powered [63]. Therefore, power effi- separation, adaptive filtering, and deep learning models to
ciency not only affects battery life but also directly impacts extract meaningful neural data from contaminated signals
wearability and biocompatibility. As such, low-power circuit [136]. During the device development phase, combining
design, efficient thermal dissipation structures, and proper EMC analysis with numerical simulations of virtual human
material selection must be thoroughly considered during models has become an essential step in verifying compatibil-
the hardware design stage. Thermal simulation tools should ity and safety in realistic tissue environments. Collectively,
be used to evaluate temperature distribution under various these layered EMC strategies are essential for improving
physiological and environmental conditions to ensure that system SNR and facilitating the transition of non-invasive
devices meet functional requirements while complying with BCIs from controlled laboratory settings to practical real-
thermal safety standards and enhancing user experience. world applications.
Integrated design is considered a key strategy for improving In non-invasive BCI systems, firmware update and roll-
power efficiency. For example, microchip-based integration back mechanisms are critical for ensuring long-term safety
of EEG and fNIRS sensors on a single platform reduces and reliability. However, these mechanisms remain under-
device size and weight, lowers power consumption, and explored. Given that BCI devices operate in close and
enhances signal quality and synchronization. sustained contact with the human body, firmware updates
Electromagnetic compatibility (EMC) presents a signifi- affect not only algorithm performance but also user safety
cant challenge throughout the signal acquisition, transmis- and ethical accountability. Without integrity verification
sion, and processing pipeline of non-invasive BCIs. The and rollback pathways, faulty updates may disrupt signal
issue arises from a fundamental contradiction: BCIs are acquisition, render devices inoperative, or introduce risks
designed to detect extremely weak neural electrophysiologi- to users [206]. Recent studies emphasize the need for digi-
cal signals (e.g., EEG) on the order of microvolts, while tally signed updates, redundant A/B partition storage, and
operating in environments that inevitably contain strong version-controlled rollback to establish a resilient firmware
internal and external electromagnetic interference. This lifecycle [207]. Real-world industry cases demonstrate that
interference includes power line radiation at 50/60 Hz and its when manufacturers discontinue software support or when

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 37 of 49 193

an update fails, users may be left with "orphaned" or non- are detected, the system can interrupt current commands
maintainable neuro-devices. Therefore, firmware systems for and roll back the classifier to a previously validated stable
BCIs must simultaneously address security, reversibility, and state. This combination of active threshold defense and pas-
long-term maintainability. sive error-triggered rollback aims to balance performance,
In non-invasive BCIs, safety failure and protection mecha- adaptability, and operational safety [209].
nisms primarily address the risk of unintended operations
caused by signal distortion, abnormal user states, or external
disturbances. Typical strategies include real-time monitoring 4 Conclusions and Perspectives
of user cognitive states (e.g., fatigue, distraction) and trig-
gering system-level interventions (e.g., halting device oper- As non-invasive BCI technology advances from labora-
ation, switching to manual mode, or activating emergency tory research toward real-world applications in clinical
shutdown) when potentially unsafe conditions are detected. rehabilitation, neural modulation, and daily life assis-
Additionally, multimodal physiological signal fusion (e.g., tance, its evolution underscores a deepening convergence
combining EEG and EOG) or robust control algorithms of neuroscience, AI, and flexible bioelectronics. Through
can enhance system fault tolerance and reduce reliance on systematic advancements in neural decoding algorithms
a single signal source, allowing basic safety functions to and innovative flexible materials, we are progressively
be maintained even when partial component failures occur. integrating high-precision neural monitoring and decod-
Regarding safety thresholds and recovery mechanisms during ing technologies seamlessly into daily life, clinical diag-
online learning, systems generally require predefined behav- nostics, and therapeutic interventions. However, achiev-
ioral boundaries or risk thresholds to constrain algorithmic ing truly sustainable, accessible, and clinically valuable
outputs, preventing learning models or adaptive control strat- applications still faces a series of challenges rooted in the
egies from entering hazardous or uncontrollable states. For intrinsic properties of neural signals, the constraints of
example, in brain-controlled robots or vehicles, physical or dynamic human-computer interaction, and the complexity
logical limits may be imposed on movement speed, steer- of long-term usage environments. To address these chal-
ing angle, or distance to obstacles. When the system detects lenges, a systematic and multidimensional breakthrough is
that its internal state or user commands are approaching or required across signal acquisition, algorithmic modeling,
exceeding these limits, recovery procedures are activated, hardware architecture, closed-loop control strategies, and
such as aborting current commands, reverting to the last evaluation frameworks.
verified safe state, or using error-related neural signals (e.g., The core performance bottleneck of non-invasive BCIs
error-related negativity) to trigger immediate system reset primarily stems from inherent biophysical constraints.
and behavioral correction. This hybrid strategy—combining When cortical neural activity passes through multiple
proactive threshold protection and reactive error-triggered tissue layers (e.g., the skull and cerebrospinal fluid), it
recovery—ensures that interactions remain within safe opera- may cause spatial blurring and amplitude attenuation of
tional bounds during dynamic learning [208]. the signals. Consequently, scalp-recorded EEG typically
In non-invasive BCIs, online learning effectively exhibits low spatial resolution, low SNR, and high suscep-
addresses the non-stationarity of EEG signals over time, but tibility to physiological artifacts and environmental noise.
it introduces key challenges related to safety thresholds and Moreover, individual variations in anatomical structure,
recovery mechanisms. Because EEG signals drift over time, physiological states (e.g., fatigue, attention), and psycho-
adaptive updates of classifiers are necessary but must avoid logical states (e.g., motivation, emotion) further exacer-
performance degradation or user confusion due to improper bate the non-stationarity and time-varying nature of neu-
adaptation. A common approach is to set a confidence ral signals. Meanwhile, practical challenges at the user
threshold: classifier parameters are updated only when the level also demand considerable attention. Firstly, there
confidence of the current prediction exceeds this thresh- are significant individual differences in neural accessi-
old, which reduces the likelihood of learning from noisy or bility. Approximately 15%–30% of users struggle to pro-
uncertain data. Meanwhile, the system must retain recovery duce consistently decodable signals, a condition termed
capability. For instance, when error-related brain potentials "BCI illiteracy." These challenges not only undermine the
 193 Page 38 of 49 Nano-Micro Lett. (2026) 18:193

generalization of decoding algorithms but also hinder the In contrast to conventional rigid electrodes, advancements
stable and reliable decoding of user intent. in flexible bioelectronics are significantly influencing the
Despite significant advances in neural decoding and sys- design principles of neural signal acquisition interfaces.
tem architecture, the large-scale translation of non-invasive Driven by innovations in materials science, nanomaterials
BCIs into clinical and everyday settings continues to face are gaining increasing attention in the development of flex-
substantial challenges. In mobile or real-world usage sce- ible electrodes. The combination of the unique electrical
narios, BCI systems confront critical issues related to the properties of nanomaterials and the excellent mechanical
stability of the biological interface, which directly impacts compliance of flexible substrates provides a solid foundation
long-term reliability and signal quality. Two primary bot- for high-performance flexible electrodes. The mechanical
tlenecks are physical degradation and relative motion at impedance matching between highly conductive flexible
the electrode-scalp interface. Firstly, interface degrada- films and biological tissues enables low interfacial imped-
tion becomes particularly pronounced during prolonged ance and high-fidelity neural signal recording. The con-
wear. Dry electrodes eliminate the need for conductive formal contact afforded by such materials helps suppress
gel, thereby reducing the risk of skin irritation or aller- motion-induced artifacts and maintains more consistent
gic reactions; however, their rigid structures may cause contact impedance. Evidence indicates that this leads to
pressure points, skin discomfort, or even micro-abrasions improvements in the SNR and stability of raw EEG signals.
during extended use. Wet electrodes, while offering supe- Compared to traditional rigid electrodes, flexible electrodes
rior signal fidelity, carry risks of skin sensitization and not only enhance signal quality but also offer greater comfort
infection. Moreover, as the conductive gel dries over time, during extended wear, facilitating long-term monitoring.
contact impedance increases, leading to progressive signal Nevertheless, even with significant advancements in mate-
instability. Semi-dry electrodes represent a compromise rials, achieving robust performance across diverse popula-
solution, slowly releasing minimal electrolytes to maintain tions still demands fundamental algorithmic breakthroughs.
conductivity while minimizing skin irritation. However, At the preprocessing stage, advanced multimodal fusion
their design requires careful consideration of material per- algorithms integrate cross-modal sensory information (e.g.,
meability and mechanical configuration to prevent uneven visual, auditory, and tactile) to enhance signal robustness
pressure distribution and residue accumulation. To ensure and discriminability. During feature extraction, methods
biocompatibility, electrode materials should be selected based on SPD manifold geometry, CSP and its variants, and
for low allergenicity, high flexibility, and smooth surface spatiotemporal filtering incorporating source imaging priors
topography. These material properties must be combined have improved feature separability and neurophysiological
with breathable designs, regular cleaning protocols, and plausibility. In classification, end-to-end deep learning mod-
personalized fitting procedures to minimize adverse skin els (e.g., CNNs, Transformer architectures, and multimodal
reactions and enhance both user comfort and system fusion networks) enable high-performance automatic fea-
reliability. ture learning and pattern recognition, significantly boost-
Secondly, dynamic artifacts present another major chal- ing decoding accuracy and generalization. Despite these
lenge. Relative movement between the electrode and scalp advances, the efficacy of such models is largely validated
during user motion not only induces mechanical fatigue under idealized experimental conditions. Their transferabil-
and potential structural damage but also introduces motion- ity, noise resilience, and long-term stability in real-world
related artifacts. Additionally, electrolytes in sweat alter the deployments remain insufficiently demonstrated.
electrochemical characteristics of the interface, causing drift Concurrently, as BCI technologies develop, ethical, soci-
and interference that further degrade signal quality. Hair etal, and safety concerns are becoming increasingly promi-
acts as a physical barrier, impeding direct contact between nent. Ultimately, translating advanced laboratory prototypes
the electrode and scalp, increasing contact impedance, and into clinically validated and commercially viable products
exacerbating artifact magnitude. Collectively, these factors requires rigorous scientific validation, comprehensive bio-
indirectly contribute to feature representation drift and dete- compatibility testing, regulatory approval, and long-term
rioration in decoding performance. efficacy assessment. Non-invasive BCIs face dual chal-
lenges in regulatory clearance and cost modeling for both

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 39 of 49 193

disposable and reusable components. From an economic asynchronous detection mechanisms, and optimizing
perspective, cost models suggest that disposable electrodes shared control strategies. However, a key engineering
increase direct expenditure but may reduce risks of cross- constraint persists: phase lag. This lag, which includes
contamination and labor costs associated with cleaning. the cumulative time delay arising from signal acquisi-
Reusable systems exhibit lower per-use costs but require tion, wireless transmission, feature extraction, decod-
investment in sterilization equipment, personnel, and main- ing decisions, and actuator response, causes the system
tenance. An optimized framework should therefore quantify output to trail behind the user’s neural intention. This
trade-offs among regulatory compliance, clinical safety, and temporal misalignment disrupts control timing, reduces
lifecycle cost. trajectory smoothness and target acquisition accuracy,
In the development and deployment of non-invasive BCIs, and significantly diminishes the user’s sense of agency,
whether as medical devices or consumer-grade products, a thereby undermining trust and sustained engagement. In
systematic checklist encompassing electrical safety, electro- high-speed motor control or high-risk tasks, even minor
magnetic compatibility (EMC), biocompatibility, software delays can lead to serious consequences. Moreover, an
lifecycle management, and cybersecurity must be strictly inherent trade-off exists between latency and reliability:
adhered to. For medical devices, patient safety and clini- complex models improve classification accuracy but often
cal reliability are paramount, necessitating compliance with incur higher computational overhead and longer response
IEC 60601 (electrical safety and EMC), ISO 10993 (bio- times, whereas simplified algorithms reduce latency at the
compatibility of electrode patches or wearable materials), expense of decoding robustness. Of equal concern is the
IEC 62304 (medical device software lifecycle), and ISO nonlinear accumulation of false positives (false triggers)
14971 (risk management). Furthermore, registration with and false negatives (missed detections) during extended
regulatory bodies such as the FDA or under the EU’s Medi- use, which increases cognitive load and may even lead to
cal Device Regulation (MDR) is mandatory, supported by user frustration or abandonment. Meanwhile, the balance
traceable technical documentation and clinical or perfor- of control authority in shared control paradigms remains
mance validation data to ensure consistent signal quality, lacking in standardized definitions. To address these chal-
algorithmic stability, and reproducibility across diverse lenges, forward models and predictive state filters should
user populations. For consumer-grade products, while full be incorporated to model user intent, generate preemptive
medical device regulations do not apply, essential require- control commands, actively compensate for cumulative
ments still hold. These include compliance with electrical processing delays, and thereby enhance the proactivity and
safety standards (e.g., IEC 62368 or IEC 61010), EMC and responsiveness of closed-loop systems. Using preregistra-
wireless regulations (e.g., FCC/CE RED), and basic safety tion or at minimum a constrained analysis plan will help
assessments of skin contact materials (referencing low-risk limit post hoc selection.
pathways in ISO 10993). Adherence to data protection laws Looking ahead, open research benchmarks should be
(e.g., GDPR) and cybersecurity regulations governing data established to promote reproducible, comparable, and veri-
encryption, privacy, and cloud transmission is equally criti- fiable progress. Pre-registered or constrained analysis plans
cal. Additionally, commercial products must avoid using are recommended to limit post hoc selection bias. In addi-
diagnostic or therapeutic terminology, instead positioning tion, the biological interpretability of BCI systems should be
their functionality as cognitive state monitoring, wellness rigorously assessed by evaluating the stability of channel or
support, or entertainment, to mitigate the risk of being clas- regional weights across repeated sessions and across differ-
sified as medical devices. Moreover, non-invasive BCIs must ent subjects. Furthermore, comprehensive, endpoint-aligned
adhere to strict disinfection and sterilization protocols in evaluation systems should be developed, integrating tech-
clinical and research settings to ensure participant safety, nical metrics with task completion time, false alarm rates,
health, and ethical protection. subjective measures such as the NASA-TLX cognitive load
Particularly crucial is that current BCI research is scale, and long-term usage rates, thereby establishing user-
undergoing a fundamental shift from "open-loop, static" centered evaluation criteria.
systems to "closed-loop, adaptive" paradigms, with a focus At the system architecture level, advancing beyond cur-
on enhancing real-time decoding capabilities, improving rent performance limitations requires the development of
 193 Page 40 of 49 Nano-Micro Lett. (2026) 18:193

specialized hardware designed for emerging computing conceptualization, and supervision. All authors have approved the
paradigms. The traditional von Neumann architecture faces final manuscript.
energy efficiency bottlenecks when processing neural signals, Declarations
whereas neuromorphic computing approaches, such as those
utilizing memristor-based in-memory computing chips com- Conflict of Interest The authors declare no interest conflict. They
have no known competing financial interests or personal relationships
bined with event-driven spiking neural networks, provide a
that could have appeared to influence the work reported in this paper.
promising solution for ultra-low-power edge computing. This
necessitates hardware-software co-design, involving cross- Open Access This article is licensed under a Creative Commons
level optimization from materials and devices to algorithms. Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format,
Such co-design ensures integrated consideration of hard-
as long as you give appropriate credit to the original author(s) and
ware physical characteristics and algorithmic computational the source, provide a link to the Creative Commons licence, and
requirements from the initial design stages, while minimizing indicate if changes were made. The images or other third party
latency and privacy risks associated with cloud transmission. material in this article are included in the article’s Creative Com-
mons licence, unless indicated otherwise in a credit line to the
In summary, the advancement of non-invasive BCIs relies material. If material is not included in the article’s Creative Com-
not only on innovations in signal processing and algorithm mons licence and your intended use is not permitted by statutory
design but also on close integration with cutting-edge flex- regulation or exceeds the permitted use, you will need to obtain
ible bioelectronics and hardware architecture. Achieving permission directly from the copyright holder. To view a copy of
this licence, visit [Link]
practical, robust, personalized, and trustworthy BCI systems
capable of transitioning from the lab to real-world deploy-
ment requires multidisciplinary collaboration and holistic
co-optimization. This process involves not only techno- References
logical innovation but also strict regulatory compliance,
1. D.J. McFarland, J.R. Wolpaw, Brain-computer interfaces for
cost-benefit analysis, and user experience improvement
communication and control. Commun. ACM 54(5), 60–66
to ensure safety, reliability, and widespread accessibility. (2011). [Link]
Simultaneously, establishing comprehensive end-to-end 2. J.R. Wolpaw, N. Birbaumer, D.J. McFarland, G. Pfurtscheller,
evaluation systems that span from signal acquisition to final T.M. Vaughan, Brain-computer interfaces for communica-
user feedback is essential for validating the effectiveness tion and control. Clin. Neurophysiol. 113(6), 767–791 (2002).
[Link]
and reliability of diverse hardware-software configurations
3. U. Chaudhary, N. Birbaumer, A. Ramos-Murguialday, Brain-
and will serve as a key focus for the future development of computer interfaces for communication and rehabilitation.
BCI technologies. Nat. Rev. Neurol. 12(9), 513–525 (2016). [Link]
1038/​nrneu​rol.​2016.​113
Acknowledgements The authors acknowledge the National 4. E.F. Chang, Brain-computer interfaces for restoring commu-
Natural Science Foundation of China for Distinguished Young nication. N. Engl. J. Med. 391(7), 654–657 (2024). [Link]
Scholars (62325403), the National Natural Science Foundation of doi.​org/​10.​1056/​nejme​24073​63
China (62504103 and 82002454), the Basic Research Program of 5. J.J. Daly, J.R. Wolpaw, Brain-computer interfaces in neu-
Jiangsu (BK20251214), the Natural Science Foundation of Jiangsu rological rehabilitation. Lancet Neurol. 7(11), 1032–1043
Province (BK20230498), the China Postdoctoral Science Founda- (2008). [Link]
tion under Grant Number 2025T180143 and 2025M770547, the
6. Y. Wang, W. Gao, S. Yang, Q. Chen, C. Ye et al., Humanoid
Medical Scientific Research Project of Jiangsu Health Commis-
intelligent display platform for audiovisual interaction and
sion (ZD2021011), and the Jiangsu Funding Program for Excellent
sound identification. Nano-Micro Lett. 15(1), 221 (2023).
Postdoctoral Talent (2024ZB427).
[Link]
7. S. Chen, M. Chen, X. Wang, X. Liu, B. Liu et al., Brain–com-
Author Contributions SW contributed to funding acquisition,
puter interfaces in 2023–2024. Brain-X 3(1), e70024 (2025).
conceptualization, and supervision. XS contributed to literature
[Link]
collection, visualization, and writing—original draft, review, edit-
ing, and response to reviewers. XS contributed to funding acqui- 8. N.S. Card, M. Wairagkar, C. Iacobacci, X. Hou, T. Singer-
sition, conceptualization, and supervision. YG contributed to the Clark et al., An accurate and rapidly calibrating speech
literature collection, editing, and supervision. ZC contributed to neuroprosthesis. N. Engl. J. Med. 391(7), 609–618 (2024).
funding acquisition and supervision. YS contributed to funding [Link]
acquisition and supervision. LY contributed to funding acquisition,

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 41 of 49 193

9. Z. Wang, N. Shi, Y. Zhang, N. Zheng, H. Li et al., Con- 22. X. Gu, Z. Cao, A. Jolfaei, P. Xu, D. Wu et al., EEG-based
formal in-ear bioelectronics for visual and auditory brain- brain-computer interfaces (BCIs): a survey of recent studies
computer interfaces. Nat. Commun. 14(1), 4213 (2023). on signal sensing technologies and computational intelligence
[Link] approaches and their applications. IEEE/ACM Trans. Com-
10. M. Mahmood, D. Mzurikwao, Y.-S. Kim, Y. Lee, S. Mishra put. Biol. Bioinform. 18(5), 1645–1666 (2021). [Link]
et al., Fully portable and wireless universal brain-machine org/​10.​1109/​TCBB.​2021.​30528​11
interfaces enabled by flexible scalp electronics and deep 23. E. Perez-Valero, M.A. Lopez-Gordo, C. Morillas, F. Pelayo,
learning algorithm. Nat. Mach. Intell. 1(9), 412–422 M.A. Vaquero-Blasco, A review of automated techniques for
(2019). [Link] assisting the early detection of Alzheimer’s disease with a
11. H. Gao, L. Luo, M. Pi, Z. Li, Q. Li et al., EEG-based voli- focus on EEG. J. Alzheimers Dis. 80(4), 1363–1376 (2021).
tional control of prosthetic legs for walking in different [Link]
terrains. IEEE Trans. Autom. Sci. Eng. 18(2), 530–540 24. Z. Gao, X. Wang, Y. Yang, Y. Li, K. Ma et al., A channel-
(2021). [Link] fused dense convolutional network for EEG-based emotion
12. S.L. Metzger, J.R. Liu, D.A. Moses, M.E. Dougherty, M.P. recognition. IEEE Trans. Cogn. Dev. Syst. 13(4), 945–954
Seaton et al., Generalizable spelling using a speech neu- (2021). [Link]
roprosthesis in an individual with severe limb and vocal 25. Z.-Z. Ma, J.-J. Wu, Z. Cao, X.-Y. Hua, M.-X. Zheng et al.,
paralysis. Nat. Commun. 13(1), 6510 (2022). [Link] Motor imagery-based brain-computer interface rehabilitation
org/​10.​1038/​s41467-​022-​33611-3 programs enhance upper extremity performance and cortical
13. A.L. Benabid, T. Costecalde, A. Eliseyev, G. Charvet, A. activation in stroke patients. J. Neuroeng. Rehabil. 21(1), 91
Verney et al., An exoskeleton controlled by an epidural (2024). [Link]
wireless brain-machine interface in a tetraplegic patient: 26. A. Biasiucci, R. Leeb, I. Iturrate, S. Perdikis, A. Al-Khodairy
a proof-of-concept demonstration. Lancet Neurol. 18(12), et al., Brain-actuated functional electrical stimulation elicits
1112–1122 (2019). [Link] lasting arm motor recovery after stroke. Nat. Commun. 9(1),
30321-7 2421 (2018). [Link]
14. F. Li, J. Gallego, N.N. Tirko, J. Greaser, D. Bashe et al., Low- 27. E.W. Sellers, D.B. Ryan, C.K. Hauser, Noninvasive brain-
intensity pulsed ultrasound stimulation (LIPUS) modulates computer interface enables communication after brainstem
microglial activation following intracortical microelectrode stroke. Sci. Transl. Med. 6(257), 257re7 (2014). [Link]
implantation. Nat. Commun. 15(1), 5512 (2024). [Link] org/​10.​1126/​scitr​anslm​ed.​30078​01
org/​10.​1038/​s41467-​024-​49709-9 28. R. Kaveh, C. Schwendeman, L. Pu, A.C. Arias, R. Muller,
15. D.V. Lam, A. Javadekar, N. Patil, M. Yu, L. Li et al., Platelets Wireless ear EEG to monitor drowsiness. Nat. Commun. 15,
and hemostatic proteins are co-localized with chronic neuro- 6520 (2024). [Link]
inflammation surrounding implanted intracortical microelec- 29. H. Si-Mohammed, J. Petit, C. Jeunet, F. Argelaguet, F. Spin-
trodes. Acta Biomater. 166, 278–290 (2023). [Link] dler et al., Towards BCI-based interfaces for augmented
10.​1016/j.​actbio.​2023.​05.​004 reality: feasibility, design and evaluation. IEEE Trans. Vis.
16. L. Du, H. Hao, Y. Ding, A. Gabros, T.C.E. Mier et al., An Comput. Graph. 26(3), 1608–1621 (2020). [Link] oi.o​ rg/1​ 0.​
implantable, wireless, battery-free system for tactile pressure 1109/​TVCG.​2018.​28737​37
sensing. Microsyst. Nanoeng. 9, 130 (2023). [Link] 30. C.-T. Lin, T.N. Do, Direct-sense brain-computer interfaces
10.​1038/​s41378-​023-​00602-3 and wearable computers. IEEE Trans. Syst. Man Cybern.
17. E.C. Gordon, A.K. Seth, Ethical considerations for the use of Syst. 51(1), 298–312 (2021). [Link]
brain-computer interfaces for cognitive enhancement. PLoS 2020.​30413​82
Biol. 22(10), e3002899 (2024). [Link] 31. N. Kaongoen, M. Yu, S. Jo, Two-factor authentication system
al.​pbio.​30028​99 using P300 response to a sequence of human photographs.
18. N. Rose, The Human Brain Project: social and ethical chal- IEEE Trans. Syst. Man Cybern. Syst. 50(3), 1178–1185
lenges. Neuron 82(6), 1212–1215 (2014). [Link] (2020). [Link]
1016/j.​neuron.​2014.​06.​001 32. H. Xue, D. Wang, M. Jin, H. Gao, X. Wang et al., Hydrogel
19. M.-M. Poo, China’s new ethical guidelines for the use of electrodes with conductive and substrate-adhesive layers for
brain-computer interfaces. Natl. Sci. Rev. 11(4), nwae154 noninvasive long-term EEG acquisition. Microsyst. Nanoeng.
(2024). [Link] 9, 79 (2023). [Link]
20. Y. Liu, Z. Zhao, M. Xu, H. Yu, Y. Zhu et al., Decoding and 33. Z. Zhang, G. Chen, S. Yang, Ensemble support vector recur-
synthesizing tonal language speech from brain activity. Sci. rent neural network for brain signal detection. IEEE Trans.
Adv. 9(23), eadh0478 (2023). [Link] Neural Netw. Learn. Syst. 33(11), 6856–6866 (2022). [Link]
adh04​78 doi.​org/​10.​1109/​TNNLS.​2021.​30837​10
21. L. Liu, B. Xi, Y. Luo, Y. Liu, P. Huang et al., Intraoperative 34. P. Sawangjai, M. Trakulruangroj, C. Boonnag, M. Piriya-
and spatiotemporal mapping of acute inflammation response jitakonkij, R.K. Tripathy et al., Eeganet: removal of ocu-
during neuroelectrode implantation. Matter 8(11), 102262 lar artifacts from the EEG signal using generative adver-
(2025). [Link] sarial networks. IEEE J. Biomed. Health Inform. 26(10),
 193 Page 42 of 49 Nano-Micro Lett. (2026) 18:193

4913–4924 (2022). [Link] oi.​o rg/​1 0.​1 109/​J BHI.​2 021.​ 47. M. Sun, W. Cui, S. Yu, H. Han, B. Hu et al., A dual-branch
31311​04 dynamic graph convolution based adaptive TransFormer
35. X. Shi, H. Wang, B. Li, Y. Qin, C. Peng et al., Fusion analy- feature fusion network for EEG emotion recognition. IEEE
sis of EEG-fNIRS multimodal brain signals: a multitask Trans. Affect. Comput. 13(4), 2218–2228 (2022). [Link]
classification algorithm incorporating spatial-temporal doi.​org/​10.​1109/​TAFFC.​2022.​31990​75
convolution and dual attention mechanisms. IEEE Trans. 48. C. Li, Z. Zhang, X. Zhang, G. Huang, Y. Liu et al., EEG-
Instrum. Meas. 74, 2506312 (2025). [Link] oi.​o rg/​1 0.​ based emotion recognition via transformer neural archi-
1109/​TIM.​2025.​35380​86 tecture search. IEEE Trans. Ind. Inform. 19(4), 6016–6025
36. S.B. Borgheai, J. McLinden, A.H. Zisk, S.I. Hosni, R.J. (2023). [Link]
Deligani et al., Enhancing communication for people in 49. S.M. Shafiul Hasan, M.R. Siddiquee, J.S. Marquez, O.
late-stage ALS using an fNIRS-based BCI system. IEEE Bai, Enhancement of movement intention detection using
Trans. Neural Syst. Rehabil. Eng. 28(5), 1198–1207 (2020). EEG signals responsive to emotional music stimulus. IEEE
[Link] Trans. Affect. Comput. 13(3), 1637–1650 (2022). [Link]
37. J. Tang, A. LeBel, S. Jain, A.G. Huth, Semantic recon- doi.​org/​10.​1109/​TAFFC.​2020.​30250​04
struction of continuous language from non-invasive brain 50. C. Wang, H. Yan, W. Huang, W. Sheng, Y. Wang et al.,
recordings. Nat. Neurosci. 26(5), 858–866 (2023). [Link] Neural encoding with unsupervised spiking convolutional
doi.​org/​10.​1038/​s41593-​023-​01304-9 neural network. Commun. Biol. 6, 880 (2023). [Link]
38. D.S. Kluger, C. Forster, O. Abbasi, N. Chalas, A. Villringer org/​10.​1038/​s42003-​023-​05257-4
et al., Modulatory dynamics of periodic and aperiodic 51. K. Kasahara, C.S. DaSalla, M. Honda, T. Hanakawa, Basal
activity in respiration-brain coupling. Nat. Commun. 14(1), ganglia-cortical connectivity underlies self-regulation of
4699 (2023). [Link] brain oscillations in humans. Commun. Biol. 5(1), 712
39. R. Zhang, S. Feng, N. Hu, S. Low, M. Li et al., Hybrid (2022). [Link]
brain-computer interface controlled soft robotic glove 52. T. Rohe, A.-C. Ehlis, U. Noppeney, The neural dynamics
for stroke rehabilitation. IEEE J. Biomed. Health Inform. of hierarchical Bayesian causal inference in multisensory
28(7), 4194–4203 (2024). [Link] oi.​o rg/​1 0.​1 109/​J BHI.​ perception. Nat. Commun. 10, 1907 (2019). [Link]
2024.​33924​12 10.​1038/​s41467-​019-​09664-2
40. Y. Zhang, C.S. Nam, G. Zhou, J. Jin, X. Wang et al., Tem- 53. X. Lin, X. Zhang, J. Chen, J. Liu, Material selection and
porally constrained sparse group spatial patterns for motor device design of scalable flexible brain-computer inter-
imagery BCI. IEEE Trans. Cybern. 49(9), 3322–3332 (2019). faces: a balance between electrical and mechanical perfor-
[Link] mance. Adv. Mater. 37(26), 2413938 (2025). [Link]
41. Z. Tang, X. Wang, J. Wu, Y. Ping, X. Guo et al., A BCI paint- org/​10.​1002/​adma.​20241​3938
ing system using a hybrid control approach based on SSVEP 54. J. Wang, T. Wang, H. Liu, K. Wang, K. Moses et al., Flex-
and P300. Comput. Biol. Med. 150, 106118 (2022). [Link] ible electrodes for brain-computer interface system. Adv.
doi.​org/​10.​1016/j.​compb​iomed.​2022.​106118 Mater. 35(47), 2211012 (2023). [Link]
42. H. Cecotti, A. Graser, Convolutional neural networks for adma.​20221​1012
P300 detection with application to brain-computer inter- 55. S. Lin, J. Jiang, K. Huang, L. Li, X. He et al., Advanced
faces. IEEE Trans. Pattern Anal. Mach. Intell. 33(3), 433–445 electrode technologies for noninvasive brain-computer
(2011). [Link] interfaces. ACS Nano 17(24), 24487–24513 (2023). [Link]
43. X. Tang, H. Shen, S. Zhao, N. Li, J. Liu, Flexible brain-com- doi.​org/​10.​1021/​acsna​no.​3c067​81
puter interfaces. Nat. Electron. 6(2), 109–118 (2023). [Link] 56. G. Balakrishnan, J. Song, C. Mou, C.J. Bettinger, Recent
doi.​org/​10.​1038/​s41928-​022-​00913-9 progress in materials chemistry to advance flexible bioel-
44. M.A. Schwemmer, N.D. Skomrock, P.B. Sederberg, J.E. Ting, ectronics in medicine. Adv. Mater. 34(10), 2106787 (2022).
G. Sharma et al., Meeting brain-computer interface user per- [Link]
formance expectations using a deep neural network decoding 57. D. Li, T.-R. Cui, J.-H. Liu, W.-C. Shao, X. Liu et al.,
framework. Nat. Med. 24(11), 1669–1676 (2018). [Link] oi.​ Motion-unrestricted dynamic electrocardiogram system
org/​10.​1038/​s41591-​018-​0171-y utilizing imperceptible electronics. Nat. Commun. 16(1),
45. M.W. Mathis, A. Perez Rotondo, E.F. Chang, A.S. Tolias, A. 3259 (2025). [Link]
Mathis, Decoding the brain: from neural representations to 58. O. Gul, M. Song, C.-Y. Gu, J. Ahn, K. Lee et al., Bioin-
mechanistic models. Cell 187(21), 5814–5832 (2024). [Link] spired interfacial engineering for highly stretchable elec-
doi.​org/​10.​1016/j.​cell.​2024.​08.​051 tronics. Nat. Commun. 16, 1337 (2025). [Link]
46. Y. Wang, M. Han, L. Jing, Q. Jia, S. Lv et al., Enhanced 1038/​s41467-​025-​56502-9
neural activity detection with microelectrode arrays modified 59. H. Cho, I. Lee, J. Jang, J.-H. Kim, H. Lee et al., Real-time
by drug-loaded calcium alginate/chitosan hydrogel. Biosens. finger motion recognition using skin-conformable electron-
Bioelectron. 267, 116837 (2025). [Link] ics. Nat. Electron. 6(8), 619–629 (2023). [Link]
bios.​2024.​116837 1038/​s41928-​023-​01012-z

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 43 of 49 193

60. P. Li, W. Sun, J. Li, J.-P. Chen, X. Wang et al., N-type semi- injury recognition. Small 20(45), 2402588 (2024). [Link]
conducting hydrogel. Science 384(6695), 557–563 (2024). doi.​org/​10.​1002/​smll.​20240​2588
[Link] 73. Y. Zhao, S. Zhao, X. Pang, A. Zhang, C. Li et al., Biomi-
61. W. Zhang, J. Liu, W. Song, J. Shan, H. Guan et al., Chemi- metic wafer-scale alignment of tellurium nanowires for high-
cal passivation and grain-boundary manipulation via in situ mobility flexible and stretchable electronics. Sci. Adv. 10(14),
cross-linking strategy for scalable flexible perovskite solar eadm9322 (2024). [Link]
cells. Sci. Adv. 11(5), eadr2290 (2025). [Link] 74. J. Zhang, T. Sun, Y. Chen, L. Liu, H. Zhao et al., Nanowires
1126/​sciadv.​adr22​90 in flexible sensors: structure is becoming a key in control-
62. E.O. Polat, G. Mercier, I. Nikitskiy, E. Puma, T. Galan et al., ling the sensing performance. Adv. Mater. Technol. 7(12),
Flexible graphene photodetectors for wearable fitness moni- 2200163 (2022). [Link]
toring. Sci. Adv. 5(9), eaaw7846 (2019). [Link] 75. Z. Zhang, Y. Wang, T. Cai, B. Wu, B. Hu et al., Recent
1126/​sciadv.​aaw78​46 progress on flexible silicon nanomembranes for advanced
63. S. Wang, J. Liu, X. Song, H. Xu, Y. Gu et al., Artificial intel- electronics and optoelectronics. Adv. Funct. Mater. 35(35),
ligence empowers solid-state batteries for material screening 2502191 (2025). [Link]
and performance evaluation. Nano-Micro Lett. 17(1), 287 76. Y. Chen, Y. Zhang, Z. Liang, Y. Cao, Z. Han et al., Flex-
(2025). [Link] ible inorganic bioelectronics. npj Flex. Electron. 4, 2 (2020).
64. L.L. Zhang, X. Zhao, M.D. Stoller, Y. Zhu, H. Ji et al., [Link]
Highly conductive and porous activated reduced graphene 77. J. Wang, Q. Zhou, A. Wang, Z. Zhu, K. Moses et al., Sponge
oxide films for high-power supercapacitors. Nano Lett. 12(4), inspired flexible, antibacterial aerogel electrode with long-
1806–1812 (2012). [Link] term high-quality electrophysiological signal recording
65. M.D. Ho, Y. Ling, L.W. Yap, Y. Wang, D. Dong et al., Perco- for human-machine interface. Adv. Funct. Mater. 34(23),
lating network of ultrathin gold nanowires and silver nanow- 2309704 (2024). [Link]
ires toward “invisible” wearable sensors for detecting emo- 78. S. Wu, H. Kan, J. Gao, W. Yue, C. Zhang et al., Convolu-
tional expression and apexcardiogram. Adv. Funct. Mater. tional neural networks-motivated high-performance multi-
27(25), 1700845 (2017). [Link] oi.o​ rg/1​ 0.1​ 002/a​ dfm.2​ 0170​ functional electronic skin for intelligent human-computer
0845 interaction. Nano Energy 122, 109313 (2024). [Link]
66. C. Xu, Y. Wang, J. Zhang, J. Wan, Z. Xiang et al., Three- org/​10.​1016/j.​nanoen.​2024.​109313
dimensional micro strain gauges as flexible, modular tactile 79. J. Huang, S. Yang, X. Tang, L. Yang, W. Chen et al., Flexible,
sensors for versatile integration with micro- and macroelec- transparent, and wafer-scale artificial synapse array based on
tronics. Sci. Adv. 10(34), eadp6094 (2024). [Link] ­TiOx/Ti3C2Tx film for neuromorphic computing. Adv. Mater.
10.​1126/​sciadv.​adp60​94 35(33), e2303737 (2023). [Link]
67. Z. Pu, X. Zhang, H. Yu, J. Tu, H. Chen et al., A thermal 20230​3737
activated and differential self-calibrated flexible epidermal 80. C. Yiu, Y. Liu, W. Park, J. Li, X. Huang et al., Skin-interfaced
biomicrofluidic device for wearable accurate blood glucose multimodal sensing and tactile feedback system as enhanced
monitoring. Sci. Adv. 7(5), eabd0199 (2021). [Link] human-machine interface for closed-loop drone control. Sci.
10.​1126/​sciadv.​abd01​99 Adv. 11(13), eadt6041 (2025). [Link] oi.o​ rg/1​ 0.1​ 126/s​ ciadv.​
68. Y. Lee, J. Kim, H. Joo, M.S. Raj, R. Ghaffari et al., Wear- adt60​41
able sensing systems with mechanically soft assemblies of 81. C. Vidaurre, B. Blankertz, Towards a cure for BCI illiteracy.
nanoscale materials. Adv. Mater. Technol. 2(9), 1700053 Brain Topogr. 23(2), 194–198 (2010). [Link]
(2017). [Link] 1007/​s10548-​009-​0121-6
69. H.-R. Lim, Y. Lee, K.A. Jones, Y.-T. Kwon, S. Kwon et al., 82. C. Rubinos, S.B. Kwon, M. Megjhani, K. Terilli, B. Wong
All-in-one, wireless, fully flexible sodium sensor system et al., Predicting shunt dependency from the effect of cer-
with integrated Au/CNT/Au nanocomposites. Sens. Actuat. ebrospinal fluid drainage on ventricular size. Neurocrit.
B Chem. 331, 129416 (2021). [Link] Care 37(3), 670–677 (2022). https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 007/​
2020.​129416 s12028-​022-​01538-8
70. H. Wang, W. Wang, J.J. Kim, C. Wang, Y. Wang et al., An 83. D. Hagemann, J. Hewig, C. Walter, E. Naumann, Skull
optical-based multipoint 3-axis pressure sensor with a flexible thickness and magnitude of EEG alpha activity. Clin. Neuro-
thin-film form. Sci. Adv. 9(36), eadi2445 (2023). [Link] physiol. 119(6), 1271–1280 (2008). [Link] oi.o​ rg/1​ 0.1​ 016/j.​
org/​10.​1126/​sciadv.​adi24​45 clinph.​2008.​02.​010
71. M. Herber, E.H. Hill, Optically-directed bubble printing of 84. M.U. Ali, A. Zafar, K.D. Kallu, H. Masood, M.M.N. Mannan
MXenes on flexible substrates toward MXene-enabled wear- et al., Correlation-filter-based channel and feature selection
able electronics and strain sensors. Nano Lett. 25(18), 7258– framework for hybrid EEG-fNIRS BCI applications. IEEE
7265 (2025). [Link] J. Biomed. Health Inform. 28(6), 3361–3370 (2024). [Link]
72. C. Yang, H. Wang, G. Zhou, H. Zhao, W. Hou et al., A multi- doi.​org/​10.​1109/​JBHI.​2023.​32945​86
modal perception-enabled flexible memristor with combined 85. D. Kollias, S. Zafeiriou, Exploiting multi-CNN features
sensing-storage-memory functions for enhanced artificial in CNN-RNN based dimensional emotion recognition on
 193 Page 44 of 49 Nano-Micro Lett. (2026) 18:193

the OMG in-the-wild dataset. IEEE Trans. Affect. Comput. Trans. Neural Netw. Learn. Syst. 34(12), 10955–10969
12(3), 595–606 (2021). [Link] oi.​o rg/​1 0.​1 109/​TAFFC.​ (2023). [Link]
2020.​30141​71 98. P. Li, C. Wang, M. Li, X. Xuan, B. Zhou et al., Flexible sil-
86. C. Li, Z. Zhang, R. Song, J. Cheng, Y. Liu et al., EEG-based ver/carbon nanotube-graphene oxide-polydimethylsiloxane
emotion recognition via neural architecture search. IEEE electrode patch for electroencephalography language. Adv.
Trans. Affect. Comput. 14(2), 957–968 (2023). [Link] Intell. Syst. 5(8), 2300018 (2023). [Link]
org/​10.​1109/​TAFFC.​2021.​31303​87 aisy.​20230​0018
87. B.J. Edelman, S. Zhang, G. Schalk, P. Brunner, G. Müller- 99. E. Yin, T. Zeyl, R. Saab, D. Hu, Z. Zhou et al., An auditory-
Putz et al., Non-invasive brain-computer interfaces: state tactile visual saccade-independent P300 brain-computer
of the art and trends. IEEE Rev. Biomed. Eng. 18, 26–49 interface. Int. J. Neural Syst. 26(1), 1650001 (2016). [Link]
(2025). [Link] doi.​org/​10.​1142/​S0129​06571​65000​15
88. X. Gao, Y. Wang, X. Chen, B. Liu, S. Gao, Brain-computer 100. X.-A. Fan, L. Bi, T. Teng, H. Ding, Y. Liu, A brain-computer
interface: a brain-in-the-loop communication system. Proc. interface-based vehicle destination selection system using
IEEE 113(5), 478–511 (2025). [Link] oi.​o rg/​1 0.​1 109/​ P300 and SSVEP signals. IEEE Trans. Intell. Transp. Syst.
JPROC.​2025.​36003​89 16(1), 274–283 (2015). [Link]
89. Y. Lu, L. Bi, H. Li, Model predictive-based shared control 23300​00
for brain-controlled driving. IEEE Trans. Intell. Transp. 101. M.-H. Lee, O.-Y. Kwon, Y.-J. Kim, H.-K. Kim, Y.-E. Lee
Syst. 21(2), 630–640 (2020). [Link] et al., EEG dataset and OpenBMI toolbox for three BCI para-
2019.​28973​56 digms: an investigation into BCI illiteracy. Gigascience 8(5),
90. E.A. Aydin, O.F. Bay, I. Guler, P300-based asynchronous giz002 (2019). [Link]
brain computer interface for environmental control system. 102. S. Mirjalili, A. Duarte, Using machine learning to simulta-
IEEE J. Biomed. Health Inform. 22(3), 653–663 (2018). neously quantify multiple cognitive components of episodic
[Link] memory. Nat. Commun. 16(1), 2856 (2025). [Link]
91. G. Ganesh, K. Nakamura, S. Saetia, A.M. Tobar, E. Yoshida 10.​1038/​s41467-​025-​58265-9
et al., Utilizing sensory prediction errors for movement 103. P.A. Kragel, M.C. Reddan, K.S. LaBar, T.D. Wager, Emotion
intention decoding: a new methodology. Sci. Adv. 4(5), schemas are embedded in the human visual system. Sci. Adv.
eaaq0183 (2018). [Link] 5(7), eaaw4358 (2019). [Link] oi.o​ rg/1​ 0.1​ 126/s​ ciadv.a​ aw43​
92. H. Wang, Y. Qi, L. Yao, Y. Wang, D. Farina et al., A 58
human-machine joint learning framework to boost endog- 104. R.M. Cichy, A. Oliva, A M/EEG-fMRI fusion primer: resolv-
enous BCI training. IEEE Trans. Neural Netw. Learn. Syst. ing human brain responses in space and time. Neuron 107(5),
35(12), 17534–17548 (2024). [Link] oi.​o rg/​1 0.​1 109/​ 772–781 (2020). [Link]
TNNLS.​2023.​33056​21 001
93. B.-Y. Tsai, S.V.S. Diddi, L.-W. Ko, S.-J. Wang, C.-Y. Chang 105. X. Ru, K. He, B. Lyu, D. Li, W. Xu et al., Multimodal neuro-
et al., Development of an adaptive artifact subspace recon- imaging with optically pumped magnetometers: a simultane-
struction based on hebbian/anti-hebbian learning networks ous MEG-EEG-fNIRS acquisition system. Neuroimage 259,
for enhancing BCI performance. IEEE Trans. Neural Netw. 119420 (2022). [Link]
Learn. Syst. 35(1), 348–361 (2024). [Link] oi.​o rg/​1 0.​ 119420
1109/​TNNLS.​2022.​31745​28 106. C.-H. Han, K.-R. Müller, H.-J. Hwang, Enhanced perfor-
94. A. Rakshit, A. Konar, A.K. Nagar, A hybrid brain-computer mance of a brain switch by simultaneous use of EEG and
interface for closed-loop position control of a robot arm. NIRS data for asynchronous brain-computer interface. IEEE
IEEE/CAA J. Autom. Sin. 7(5), 1344–1360 (2020). [Link] Trans. Neural Syst. Rehabil. Eng. 28(10), 2102–2112 (2020).
doi.​org/​10.​1109/​JAS.​2020.​10033​36 [Link]
95. S. Pancholi, A. Giri, A. Jain, L. Kumar, S. Roy, Source 107. Y.-C. Jiang, R. Ma, S. Qi, S. Ge, Z. Sun et al., Characteriza-
aware deep learning framework for hand kinematic recon- tion of bimanual cyclical tasks from single-trial EEG-fNIRS
struction using EEG signal. IEEE Trans. Cybern. 53(7), measurements. IEEE Trans. Neural Syst. Rehabil. Eng. 30,
4094–4106 (2023). [Link] oi.​o rg/​1 0.​1 109/​TCYB.​2022.​ 146–156 (2022). [Link] oi.o​ rg/1​ 0.1​ 109/T
​ NSRE.2​ 022.3​ 1442​
31666​04 16
96. C. Ju, C. Guan, Graph neural networks on SPD manifolds 108. M.A. Pisauro, E. Fouragnan, C. Retzler, M.G. Philiastides,
for motor imagery classification: a perspective from the Neural correlates of evidence accumulation during value-
time-frequency analysis. IEEE Trans. Neural Netw. Learn. based decisions revealed via simultaneous EEG-fMRI. Nat.
Syst. 35(12), 17701–17715 (2024). [Link] Commun. 8, 15808 (2017). [Link]
TNNLS.​2023.​33074​70 s15808
97. C. Ju, C. Guan, Tensor-CSPNet: a novel geometric deep 109. D. Ji, Y. Huang, Z. Chen, X. Zhou, J. Wang et al., Enhanced
learning framework for motor imagery classification. IEEE spatial division multiple access BCI performance via incor-
porating MEG with EEG. IEEE Trans. Neural Syst. Rehabil.

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 45 of 49 193

Eng. 33, 1202–1211 (2025). [Link] oi.o​ rg/1​ 0.1​ 109/T

​ NSRE.​ 121. D. Forenzo, H. Zhu, B. He, A continuous pursuit dataset
2025.​35506​53 for online deep learning-based EEG brain-computer inter-
110. M. Klumpp, L. Embray, F. Heimburg, A.L. Alves Dias, J. face. Sci. Data 11(1), 1256 (2024). [Link]
Simon et al., Syntalos: a software for precise synchronization s41597-​024-​04090-6
of simultaneous multi-modal data acquisition and closed-loop 122. Y. Li, L. Fan, H. Shen, D. Hu, HR-SNN: an end-to-end
interventions. Nat. Commun. 16(1), 708 (2025). [Link] spiking neural network for four-class classification motor
org/​10.​1038/​s41467-​025-​56081-9 imagery brain–computer interface. IEEE Trans. Cogn. Dev.
111. W. Li, C. Gao, Z. Li, Y. Diao, J. Li et al., BrainFusion: a low- Syst. 16(6), 1955–1968 (2024). [Link] oi.​o rg/​1 0.​1 109/​
code, reproducible, and deployable software framework for TCDS.​2024.​33954​43
multimodal brain-computer interface and brain-body interac- 123. C. Li, Z. Gu, A Riemannian convolutional neural network
tion research. Adv. Sci. 12(32), e17408 (2025). [Link] for EEG-based motor imagery decoding. Neurocomput-
org/​10.​1002/​advs.​20241​7408 ing 639, 130206 (2025). [Link]
112. G. Castrillon, N. Sollmann, K. Kurcyus, A. Razi, S.M. Krieg 2025.​130206
et al., The physiological effects of noninvasive brain stimula- 124. S. Huang, Q. Wei, A deep learning model combining con-
tion fundamentally differ across the human cortex. Sci. Adv. volutional neural networks and a selective kernel mecha-
6(5), eaay2739 (2020). [Link] nism for SSVEP-Based BCIs. Comput. Biol. Med. 196,
39 110691 (2025). [Link] oi.​o rg/​1 0.​1 016/j.​c ompb​i omed.​
113. P. Georgieva, N. Bouaynaya, F. Silva, L. Mihaylova, L.C. 2025.​110691
Jain, A beamformer-particle filter framework for localization 125. F. Fahimi, S. Dosen, K.K. Ang, N. Mrachacz-Kersting, C.
of correlated EEG sources. IEEE J. Biomed. Health Inform. Guan, Generative adversarial networks-based data augmenta-
20(3), 880–892 (2016). [Link] tion for brain-computer interface. IEEE Trans. Neural Netw.
24137​52 Learn. Syst. 32(9), 4039–4051 (2021). [Link]
114. S. Iwama, M. Takemi, R. Eguchi, R. Hirose, M. Morishige 1109/​TNNLS.​2020.​30166​66
et al., Two common issues in synchronized multimodal 126. J. Zhang, J. Fang, S. Liu, D. Liu, H. Wu et al., Toward cross-
recordings with EEG: jitter and latency. Neurosci. Res. 203, brain-computer interface: a prototype-supervised adversarial
1–7 (2024). [Link] transfer learning approach with multiple sources. IEEE Trans.
115. E. Boto, Z.A. Seedat, N. Holmes, J. Leggett, R.M. Hill et al., Instrum. Meas. 73, 2528113 (2024). [Link]
Wearable neuroimaging: combining and contrasting magne- TIM.​2024.​34515​93
toencephalography and electroencephalography. Neuroimage 127. A. Momeni, B. Rahmani, M. Malléjac, P. Del Hougne, R.
201, 116099 (2019). [Link] Fleury, Backpropagation-free training of deep physical neural
2019.​116099 networks. Science 382(6676), 1297–1303 (2023). [Link] oi.​
116. L. Qiu, W. Feng, L. Zhong, X. Song, Z. Ying et al., Mutual org/​10.​1126/​scien​ce.​adi84​74
information learning-based end-to-end fusion network for 128. B. Yuen, X. Dong, T. Lu, A 3D ray traced biological neural
hybrid EEG-fNIRS brain-computer interface. IEEE Trans. network learning model. Nat. Commun. 15(1), 4693 (2024).
Instrum. Meas. 74, 6510113 (2025). [Link] [Link]
TIM.​2025.​36049​29 129. W. Fang, Y. Chen, J. Ding, Z. Yu, T. Masquelier et al., Spik-
117. N.J. Sexton, B.C. Love, Reassessing hierarchical correspond- ingJelly: an open-source machine learning infrastructure plat-
ences between brain and deep networks through direct inter- form for spike-based intelligence. Sci. Adv. 9(40), eadi1480
face. Sci. Adv. 8(28), eabm2219 (2022). [Link] (2023). [Link]
1126/​sciadv.​abm22​19 130. T. Zhang, X. Cheng, S. Jia, M.-M. Poo, Y. Zeng et al., Self-
118. G.B. Kshirsagar, N.D. Londhe, DS-P3SNet: an efficient clas- backpropagation of synaptic modifications elevates the effi-
sification approach for Devanagari script-based P300 speller ciency of spiking and artificial neural networks. Sci. Adv.
using compact channelwise convolution and knowledge dis- 7(43), eabh0146 (2021). [Link] oi.o​ rg/1​ 0.1​ 126/s​ ciadv.a​ bh01​
tillation. IEEE Trans. Syst. Man Cybern. Syst. 52(12), 7431– 46
7443 (2022). [Link] 131. G. Amiri, V. Shalchyan, Decoding muscle activity via CNN-
119. Y. Ding, N. Robinson, C. Tong, Q. Zeng, C. Guan, LGGNet: LSTM from 3D spatiotemporal EEG. Comput. Methods
learning from local-global-graph representations for brain- Progr. Biomed. 271, 108983 (2025). [Link] oi.o​ rg/1​ 0.1​ 016/j.​
computer interface. IEEE Trans. Neural Netw. Learn. Syst. cmpb.​2025.​108983
35(7), 9773–9786 (2024). [Link] 132. Z. Khademi, F. Ebrahimi, H.M. Kordy, A transfer learning-
2023.​32366​35 based CNN and LSTM hybrid deep learning model to clas-
120. E. Jeon, W. Ko, J.S. Yoon, H.-I. Suk, Mutual information- sify motor imagery EEG signals. Comput. Biol. Med. 143,
driven subject-invariant and class-relevant deep representa- 105288 (2022). [Link]
tion learning in BCI. IEEE Trans. Neural Netw. Learn. Syst. 105288
34(2), 739–749 (2023). [Link] 133. Z. An, F. Wang, Y. Wen, F. Hu, S. Han, A real-time CNN–
2021.​31005​83 BiLSTM-based classifier for patient-centered AR-SSVEP
active rehabilitation exoskeleton system. Expert Syst. Appl.
 193 Page 46 of 49 Nano-Micro Lett. (2026) 18:193

255, 124706 (2024). [Link] 146. R.E. Alcaide-Aguirre, S.A. Warschausky, D. Brown, A. Aref,
124706 J.E. Huggins, Asynchronous brain-computer interface for
134. J.-H. Jeong, J.-H. Cho, B.-H. Lee, S.-W. Lee, Real-time deep cognitive assessment in people with cerebral palsy. J. Neural
neurolinguistic learning enhances noninvasive neural lan- Eng. 14(6), 066001 (2017). [Link]
guage decoding for brain–machine interaction. IEEE Trans. 2552/​aa7fc4
Cybern. 53(12), 7469–7482 (2023). [Link] 147. X. Zhang, W. Hou, Y. Wang, N. Jiang, Low-latency asyn-
TCYB.​2022.​32116​94 chronous BCI by detecting the motion onset VEP induced
135. M. Jiao, S. Yang, X. Xian, N. Fotedar, F. Liu, Multi-modal by the action observation stimulus in senior. IEEE Trans.
electrophysiological source imaging with attention neural Instrum. Meas. 72, 2528610 (2023). [Link]
networks based on deep fusion of EEG and MEG. IEEE TIM.​2023.​33207​59
Trans. Neural Syst. Rehabil. Eng. 32, 2492–2502 (2024). 148. Z. Liu, J. Mei, J. Tang, M. Xu, B. Gao et al., A memristor-
[Link] based adaptive neuromorphic decoder for brain-computer
136. D. Lahat, T. Adali, C. Jutten, Multimodal data fusion: an interfaces. Nat. Electron. 8(4), 362–372 (2025). [Link]
overview of methods, challenges, and prospects. Proc. IEEE org/​10.​1038/​s41928-​025-​01340-2
103(9), 1449–1477 (2015). [Link] 149. J.Y. Lee, S. Lee, A. Mishra, X. Yan, B. McMahan et al.,
2015.​24606​97 Brain-computer interface control with artificial intelligence
137. H. Sun, H. Tian, Y. Hu, Y. Cui, X. Chen et al., Bio-plausible copilots. Nat. Mach. Intell. 7(9), 1510–1523 (2025). [Link]
multimodal learning with emerging neuromorphic devices. doi.​org/​10.​1038/​s42256-​025-​01090-y
Adv. Sci. 11(45), 2406242 (2024). [Link] 150. Y. Ding, C. Udompanyawit, Y. Zhang, B. He, EEG-based
advs.​20240​6242 brain-computer interface enables real-time robotic hand
138. J. Wang, L. Yao, Y. Wang, Enhanced online continuous brain- control at individual finger level. Nat. Commun. 16(1), 5401
control by deep learning-based EEG decoding. IEEE Trans. (2025). [Link]
Neural Syst. Rehabil. Eng. 33, 2834–2846 (2025). [Link] oi.​ 151. Y. Dong, W. An, Z. Wang, D. Zhang, An artificial intelli-
org/​10.​1109/​TNSRE.​2025.​35912​54 gence-assisted flexible and wearable mechanoluminescent
139. K. Bhadra, A.-L. Giraud, S. Marchesotti, Learning to oper- strain sensor system. Nano-Micro Lett. 17(1), 62 (2024).
ate an imagined speech Brain-Computer Interface involves [Link]
the spatial and frequency tuning of neural activity. Com- 152. S. Huang, B. Zhang, Z. Shao, L. He, Q. Zhang et al., Ultra-
mun. Biol. 8(1), 271 (2025). https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 038/​ miniaturized stretchable strain sensors based on single sili-
s42003-​025-​07464-7 con nanowires for imperceptible electronic skins. Nano Lett.
140. I. Batzianoulis, F. Iwane, S. Wei, C.G.P.R. Correia, R. 20(4), 2478–2485 (2020). [Link]
Chavarriaga et al., Customizing skills for assistive robotic ett.​9b052​17
manipulators, an inverse reinforcement learning approach 153. Z. Liu, J. Xu, D. Chen, G. Shen, Flexible electronics based
with error-related potentials. Commun. Biol. 4(1), 1406 on inorganic nanowires. Chem. Soc. Rev. 44(1), 161–192
(2021). [Link] (2015). [Link]
141. X. Deng, Z.L. Yu, C. Lin, Z. Gu, Y. Li, Self-adaptive shared 154. Y. Guo, K. Li, W. Yue, N.-Y. Kim, Y. Li et al., A rapid adap-
control with brain state evaluation network for human- tation approach for dynamic air-writing recognition using
wheelchair cooperation. J. Neural Eng. 17(4), 045005 (2020). wearable wristbands with self-supervised contrastive learn-
[Link] ing. Nano-Micro Lett. 17(1), 41 (2024). [Link]
142. Y. Zhou, T. Yu, W. Gao, W. Huang, Z. Lu et al., Shared three- 1007/​s40820-​024-​01545-8
dimensional robotic arm control based on asynchronous BCI 155. B. Wen, J.E. Sader, J.J. Boland, Mechanical properties of
and computer vision. IEEE Trans. Neural Syst. Rehabil. Eng. ZnO nanowires. Phys. Rev. Lett. 101(17), 175502 (2008).
31, 3163–3175 (2023). [Link] oi.o​ rg/1​ 0.1​ 109/T
​ NSRE.2​ 023.​ [Link]
32993​50 156. W. Liu, N. Liu, J. Sun, P.-C. Hsu, Y. Li et al., Ionic con-
143. L. Hu, J. Zhu, S. Chen, Y. Zhou, Z. Song et al., A wearable ductivity enhancement of polymer electrolytes with ceramic
asynchronous brain-computer interface based on EEG-EOG nanowire fillers. Nano Lett. 15(4), 2740–2745 (2015). [Link]
signals with fewer channels. IEEE Trans. Biomed. Eng. doi.​org/​10.​1021/​acs.​nanol​ett.​5b006​00
71(2), 504–513 (2024). [Link] 157. J.W. Borchert, U. Zschieschang, F. Letzkus, M. Giorgio, R.T.
33083​71 Weitz et al., Flexible low-voltage high-frequency organic
144. F. Aloise, P. Aricò, F. Schettini, S. Salinari, D. Mattia et al., thin-film transistors. Sci. Adv. 6(21), eaaz5156 (2020).
Asynchronous gaze-independent event-related potential- [Link]
based brain-computer interface. Artif. Intell. Med. 59(2), 158. H. An, T. Habib, S. Shah, H. Gao, M. Radovic et al., Surface-
61–69 (2013). [Link] agnostic highly stretchable and bendable conductive MXene
145. G.R. Muller-Putz, G. Pfurtscheller, Control of an electrical multilayers. Sci. Adv. 4(3), eaaq0118 (2018). [Link]
prosthesis with an SSVEP-based BCI. IEEE Trans. Biomed. 10.​1126/​sciadv.​aaq01​18
Eng. 55(1), 361–364 (2008). [Link] 159. Z. Ren, S. Deng, J. Shao, Y. Si, C. Zhou et al., Ultrahigh-
2007.​897815 power-density flexible piezoelectric energy harvester based

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 47 of 49 193

on freestanding ferroelectric oxide thin films. Nat. Commun. 172. S. Zhang, T. Zhang, Z. Liu, J. Wang, J. Xu et al., Flexible
16, 3192 (2025). [Link] oi.o​ rg/1​ 0.1​ 038/s​ 41467-0​ 25-5​ 8386-1 and robust 3D a-SiGe radial junction near-infrared pho-
160. D. Ma, M. Ji, H. Yi, Q. Wang, F. Fan et al., Pushing the thin- todetectors for rapid sphygmic signal monitoring. Adv.
ness limit of silver films for flexible optoelectronic devices Funct. Mater. 32(2), 2107040 (2022). [Link]
via ion-beam thinning-back process. Nat. Commun. 15(1), 1002/​adfm.​20210​7040
2248 (2024). [Link] 173. H. Ma, J. Xu, K. Chen, L. Yu, Synergetic effect in rolling
161. X. Wang, G. Bai, J. Liang, Q. Xie, Z. Chen et al., Gusta- GaIn alloy droplets enables ultralow temperature growth of
tory interface for operative assessment and taste decoding silicon nanowires at 70 °C on plastics. Nanoscale 12(16),
in patients with tongue cancer. Nat. Commun. 15(1), 8967 8949–8957 (2020). [Link]
(2024). [Link] 174. X. Song, Y. Gu, S. Wang, J. Wang, L. Yu, Nanowire-based
162. D. Kim, W.K. Jung, S. Lee, Single-crystalline-level properties flexible sensors for wearable electronics, brain-computer
of ultrathin ­SrRuO3 flexible membranes with wide and clean interfaces, and artificial skins. Electron 3(2), e77 (2025).
surface. NPJ Flex. Electron. 6, 24 (2022). [Link] [Link]
1038/​s41528-​022-​00155-x 175. X. Song, L. Wu, Y. Liang, Z. Liu, J. Wang et al., High-
163. J. Jin, Z. Zhu, Y. Ming, Y. Zhou, J. Shang et al., Sponta- performance transparent silicon nanowire thin film transis-
neous bifacial capping of perovskite film for efficient and tors integrated on glass substrates via a room temperature
mechanically stable flexible solar cell. Nat. Commun. 16(1), solution passivation. Adv. Electron. Mater. 9(4), 2201236
90 (2025). [Link] (2023). [Link]
164. B. Sun, R. Xu, X. Han, J. Xu, W. Hong et al., Ultra-high tem- 176. Z. Xue, M. Sun, T. Dong, Z. Tang, Y. Zhao et al., Deter-
perature tolerant flexible transparent electrode with embed- ministic line-shape programming of silicon nanowires for
ded silver nanowires bundle micromesh for electrical heater. extremely stretchable springs and electronics. Nano Lett.
NPJ Flex. Electron. 6, 48 (2022). [Link] 17(12), 7638–7646 (2017). [Link] oi.​o rg/​1 0.​1 021/​a cs.​
s41528-​022-​00182-8 nanol​ett.​7b036​58
165. J. Liu, C. Jiang, Q. Yu, Y. Ni, C. Yu et al., Multidimensional 177. J. Yan, Y. Zhang, Z. Liu, J. Wang, J. Xu et al., Ultracom-
free shape-morphing flexible neuromorphic devices with reg- pact single-nanowire-morphed grippers driven by vectorial
ulation at arbitrary points. Nat. Commun. 16(1), 756 (2025). Lorentz forces for dexterous robotic manipulations. Nat.
[Link] Commun. 14(1), 3786 (2023). [Link] oi.​o rg/​1 0.​1 038/​
166. K.L. Zhou, C.B. Han, C.F. Li, J. Jiu, Y. Yang et al., Highly s41467-​023-​39524-z
stable transparent conductive electrodes based on silver- 178. R. Yuan, W. Qian, Z. Liu, J. Wang, J. Xu et al., Designable
platinum alloy-walled hollow nanowires for optoelectronic integration of silicide nanowire springs as ultra-compact
devices. ACS Appl. Mater. Interfaces 10(42), 36128–36135 and stretchable electronic interconnections. Small 18(6),
(2018). [Link] 2104690 (2022). [Link]
167. D. Kim, J. Kwon, J. Jung, K. Kim, H. Lee et al., A transpar- 179. L. Wu, Z. Hu, L. Liang, R. Hu, J. Wang et al., Step-necking
ent and flexible capacitive-force touch pad from high-aspect- growth of silicon nanowire channels for high performance
ratio copper nanowires with enhanced oxidation resistance field effect transistors. Nat. Commun. 16(1), 965 (2025).
for applications in wearable electronics. Small Meth. 2(7), [Link]
1800077 (2018). [Link] 180. H. Ma, R. Yuan, J. Wang, Y. Shi, J. Xu et al., Cylindrical
168. L. Li, S. Zhao, W. Ran, Z. Li, Y. Yan et al., Dual sens- line-feeding growth of free-standing silicon nanohelices as
ing signal decoupling based on tellurium anisotropy for elastic springs and resonators. Nano Lett. 20(7), 5072–5080
VR interaction and neuro-reflex system application. Nat. (2020). [Link]
Commun. 13(1), 5975 (2022). https:// ​ d oi. ​ o rg/ ​ 1 0.​1 038/​ 181. H. Hu, S. Wang, S. Wang, G. Liu, T. Cao et al., Aligned
s41467-​022-​33716-9 silver nanowires enabled highly stretchable and transparent
169. X. Karagiorgis, N.M. Nair, S. Sandhu, A.S. Dahiya, P.J. Ska- electrodes with unusual conductive property. Adv. Funct.
bara et al., Fully degradable, transparent, and flexible pho- Mater. 29(33), 1902922 (2019). [Link]
todetectors using ZnO nanowires and PEDOT: PSS based adfm.​20190​2922
nanofibres. NPJ Flex. Electron. 9(1), 22 (2025). [Link] 182. B. Hwang, H.-A.-S. Shin, T. Kim, Y.-C. Joo, S.M. Han,
org/​10.​1038/​s41528-​025-​00385-9 Highly reliable Ag nanowire flexible transparent electrode
170. X. Li, S. Xing, Y. Li, G. Wang, X. Gao et al., Colloidal syn- with mechanically welded junctions. Small 10(16), 3397–
thesis of sub-1-nm PbSe nanowires via cation exchange for 3404 (2014). [Link]
high-performance near-infrared self-powered photoelectro- 183. V.R. Feig, H. Tran, M. Lee, Z. Bao, Mechanically tunable
chemical-type photodetectors. Adv. Sci. 12(26), 2501993 conductive interpenetrating network hydrogels that mimic
(2025). [Link] the elastic moduli of biological tissue. Nat. Commun. 9(1),
171. S. Wang, X. Song, J. Xu, J. Wang, L. Yu, Flexible silicon 2740 (2018). [Link]
for high-performance photovoltaics, photodetectors and 184. H. He, L. Zhang, X. Guan, H. Cheng, X. Liu et al., Biocom-
bio-interfaced electronics. Mater. Horiz. 12(4), 1106–1132 patible conductive polymers with high conductivity and
(2025). [Link] high stretchability. ACS Appl. Mater. Interfaces 11(29),
 193 Page 48 of 49 Nano-Micro Lett. (2026) 18:193

26185–26193 (2019). [Link] 196. L.M. Ferrari, U. Ismailov, J.-M. Badier, F. Greco, E.
25 Ismailova, Conducting polymer tattoo electrodes in clinical
185. J. Zhou, H. Liu, Y. Sun, C. Wang, K. Chen, Self-healing electro- and magneto-encephalography. NPJ Flex. Electron.
titanium dioxide nanocapsules-graphene/multi-branched 4, 4 (2020). [Link]
polyurethane hybrid flexible film with multifunctional prop- 197. W. Wang, H. Zhou, Z. Xu, Z. Li, L. Zhang et al., Flexible
erties toward wearable electronics. Adv. Funct. Mater. 31(29), conformally bioadhesive MXene hydrogel electronics for
2011133 (2021). [Link] machine learning-facilitated human-interactive sensing.
186. M. Tavakoli, M.H. Malakooti, H. Paisana, Y. Ohm, D. Green Adv. Mater. 36(31), 2401035 (2024). [Link] oi.o​ rg/1​ 0.1​ 002/​
Marques et al., EGaIn-assisted room-temperature sintering of adma.​20240​1035
silver nanoparticles for stretchable, inkjet-printed, thin-film 198. N. Driscoll, B. Erickson, B.B. Murphy, A.G. Richardson, G.
electronics. Adv. Mater. 30(29), 1801852 (2018). [Link] Robbins et al., MXene-infused bioelectronic interfaces for
org/​10.​1002/​adma.​20180​1852 multiscale electrophysiology and stimulation. Sci. Transl.
187. S.-H. Sunwoo, S.I. Han, D. Jung, M. Kim, S. Nam et al., Med. 13(612), eabf8629 (2021). [Link]
Stretchable low-impedance conductor with Ag-Au-Pt core- anslm​ed.​abf86​29
shell-shell nanowires and in situ formed Pt nanoparticles for 199. S. Lin, J. Liu, W. Li, D. Wang, Y. Huang et al., A flexible,
wearable and implantable device. ACS Nano 17(8), 7550– robust, and gel-free electroencephalogram electrode for non-
7561 (2023). [Link] invasive brain-computer interfaces. Nano Lett. 19(10), 6853–
188. Y. Hong, J. Wang, J. Li, Z. Xu, X. Yang et al., Enhancing 6861 (2019). [Link]
non-invasive brain stimulation with non-invasively delivered 200. G. Yang, K. Zhu, W. Guo, D. Wu, X. Quan et al., Adhesive
nanoparticles for improving stroke recovery. Mater. Today and hydrophobic bilayer hydrogel enabled on-skin biosensors
Chem. 26, 101104 (2022). [Link] for high-fidelity classification of human emotion. Adv. Funct.
2022.​101104 Mater. 32(29), 2200457 (2022). [Link] oi.o​ rg/1​ 0.1​ 002/a​ dfm.​
189. Q. Han, C. Zhang, T. Guo, Y. Tian, W. Song et al., Hydrogel 20220​0457
nanoarchitectonics of a flexible and self-adhesive electrode 201. M. Mahmood, N. Kim, M. Mahmood, H. Kim, H. Kim et al.,
for long-term wireless electroencephalogram recording and VR-enabled portable brain-computer interfaces via wire-
high-accuracy sustained attention evaluation. Adv. Mater. less soft bioelectronics. Biosens. Bioelectron. 210, 114333
35(12), 2209606 (2023). [Link] oi.o​ rg/1​ 0.1​ 002/a​ dma.2​ 0220​ (2022). [Link]
9606 202. C. Brunner, R. Scherer, B. Graimann, G. Supp, G. Pfurtschel-
190. K. Liu, B. Ouyang, X. Guo, Y. Guo, Y. Liu, Advances in flex- ler, Online control of a brain-computer interface using phase
ible organic field-effect transistors and their applications for synchronization. IEEE Trans. Biomed. Eng. 53(12), 2501–
flexible electronics. NPJ Flex. Electron. 6, 1 (2022). [Link] 2506 (2006). [Link]
doi.​org/​10.​1038/​s41528-​022-​00133-3 203. X. He, M. Wu, H. Li, S. Liu, B. Liu et al., Real-time regula-
191. Y. Li, Y. Gu, J. Teng, S. Zheng, Y. Pang et al., Advancing tion of room temperature based on individual thermal sen-
EEG-based brain-computer interface technology via PEDOT: sation using an online brain-computer interface. Indoor Air
PSS electrodes. Matter 7(9), 2859–2895 (2024). [Link] 32(9), e13106 (2022). [Link]
org/​10.​1016/j.​matt.​2024.​05.​023 204. C. Serrano-Amenos, P. Heydari, C.Y. Liu, A.H. Do, Z.
192. X. Song, T. Zhang, L. Wu, R. Hu, W. Qian et al., Highly Nenadic, Power budget of a skull unit in a fully-implantable
stretchable high-performance silicon nanowire field effect brain-computer interface: bio-heat model. IEEE Trans. Neu-
transistors integrated on elastomer substrates. Adv. Sci. 9(9), ral Syst. Rehabil. Eng. 31, 4029–4039 (2023). [Link] oi.o​ rg/​
2105623 (2022). [Link] 10.​1109/​TNSRE.​2023.​33239​16
193. X. Song, Y. Gu, S. Wang, J. Fan, J. An et al., Scalable integra- 205. T.S. Ibrahim, D. Abraham, R.L. Rennaker, Electromag-
tion of high sensitivity strain sensors based on silicon nanow- netic power absorption and temperature changes due
ire spring array directly grown on flexible polyimide films. to brain machine interface operation. Ann. Biomed.
Nano Lett. 25(6), 2290–2297 (2025). [Link] oi.o​ rg/1​ 0.1​ 021/​ Eng. 35(5), 825–834 (2007). [Link] oi.​o rg/​1 0.​1 007/​
acs.​nanol​ett.​4c055​53 s10439-​007-​9264-3
194. C. Shi, J. Jiang, C. Li, C. Chen, W. Jian et al., Precision- 206. S.L. Bernal, A.H. Celdrán, G.M. Pérez, M.T. Barros, S.
induced localized molten liquid metal stamps for damage- Balasubramaniam, Security in brain-computer interfaces:
free transfer printing of ultrathin membranes and 3D objects. state-of-the-art, opportunities, and future challenges. ACM
Nat. Commun. 15, 8839 (2024). [Link] Comput. Surv. 54(1), 1–35 (2022). [Link]
s41467-​024-​53184-7 34273​76
195. Y. Park, C.K. Franz, H. Ryu, H. Luan, K.Y. Cotton et al., 207. T. Schroder, R. Sirbu, S. Park, J. Morley, S. Street et al.,
Three-dimensional, multifunctional neural interfaces for cor- Cyber risks to next-gen brain-computer interfaces: analysis
tical spheroids and engineered assembloids. Sci. Adv. 7(12), and recommendations. Neuroethics 18(2), 34 (2025). [Link]
eabf9153 (2021). [Link] doi.​org/​10.​1007/​s12152-​025-​09607-3

© The authors [Link]

Nano-Micro Lett. (2026) 18:193 Page 49 of 49 193

208. F. Brocal, Brain-computer interfaces in safety and security Syst. Rehabil. Eng. 14(1), 30–37 (2006). [Link]
fields: risks and applications. Saf. Sci. 160, 106051 (2023). 1109/​TNSRE.​2005.​863842
[Link]
209. G.R. Muller-Putz, R. Scherer, C. Neuper, G. Pfurtscheller, Publisher’s Note Springer Nature remains neutral with regard to
Steady-state somatosensory evoked potentials: suitable brain jurisdictional claims in published maps and institutional affiliations.
signals for brain-computer interfaces? IEEE Trans. Neural