Convolutional Neural Network driven

Electroencephalogram Characterization for Robust

and Efficient Schizophrenia Diagnosis
Haseeb Iqbal∗ Salaar Khan† Anoosha Tahir‡ Hamza Ramzan§
∗ BiomedicalEngineering Department, University of Engineering & Technology, Pakistan
† Schoolof Science & Engineering, Lahore University of Management Sciences, Pakistan
‡ Buch International Hospital, Multan, Pakistan
§ Computer Science Department, University of Engineering & Technology, Pakistan

Abstract—Complex psychiatric disorders such as Schizophre- tasks [4]. Cyclical degradation results from the disease’s slow
nia show symptoms like disorganized thinking and hallucinations, penetration and disruption of the brain’s delicate neuronal
typically rooted in abnormal brain activity. Early detection of circuitry [5]. For schizophrenia to be effectively treated and
such abnormal brain activity significantly improves diagnosis and
subsequent treatment. To that end, Electroencephalogram (EEG), its negative effects to be minimized, a prompt diagnosis is
a non-invasive technique that records electrical activity in the necessary. The time it takes from symptoms to show to the
brain is often used to record information about these cerebral start of the treatment greatly affects the prognosis later on,
anomalies. EEG data can be analyzed using various techniques which highlights the significance of early intervention [6].
to facilitate Schizophrenia diagnosis. In that respect, traditional Initiating treatment during a crucial window of time allows for
machine learning techniques have been extensively used. How-
ever, traditional machine learning techniques necessitate pre- preventing irreversible brain damage and preserving cognitive
processing and extensive feature extraction, which can often function. Unfortunately, the sensitivity and objectivity required
considerably vary from one dataset to another. In contrast, deep for early detection are often lacking in current diagnostic
learning techniques are better suited to classify between normal techniques, which rely on subjective clinical assessments. This
and schizophrenic cases, owing to their sophisticated network causes a significant delay in receiving necessary treatment,
architecture and their ability to resolve complex data. In this
work, we leverage a convolutional neural network (CNN) to char- which worsens symptoms and impairs long-term prognosis [7].
acterize EEG data from two cohorts to facilitate schizophrenia The diagnosis of schizophrenia has traditionally been made
diagnosis. We report a 96.4% accuracy of classification with fewer through clinical interviews and observational assessments that
network parameters which beats the current gold standard and are, by nature, subjective and susceptible to inter-rater vari-
shows promise for effective and robust detection and subsequent ability. Most of these methodologies deal with the symptoma-
Schizophrenia diagnosis based on EEG data.
Index Terms—Schizophrenia, Diagnosis, CNN, EEG, Neural tology and patient history, which causes a problem because
Network schizophrenia can be very subtle and tends to produce episodic
symptoms, delaying its detection many times. The opposite
I. I NTRODUCTION change occurs with the arrival of Artificial Intelligence (AI)
Schizophrenia is a chronic neuropsychiatric disorder charac- in diagnostic methodologies. AI uses large-scale dataset anal-
terized by disruptions in thought processes, perceptions, emo- ysis, such as EEG signals, in applying complex algorithms
tional responsiveness, and social interactions. Disturbances in and machine learning models to discern subtle patterns and
speech, behavior, and cognition are hallmarks of schizophre- biomarkers that are reflective of schizophrenia. Thus, when
nia, a severe and long-lasting neurological disorder [1] [2]. compared to traditional methods, AI-enabled diagnostics are
Based on estimates from National Institute of Mental Health, without human cognitive biasing and thus can process time
2.4 million individuals in the US suffer from schizophrenia, series EEG signals to generate a comprehensive diagnostic
making it a considerable contributor to the total illness burden. profile. This would not only improve diagnostic accuracy but
In addition, the World Health Organization states that about also enable disease detection at very early stages, personalized
21 million people suffer from it, globally. This severe mental treatment planning, and continuous monitoring of disease
illness is typified by psychotic symptoms that are either progression, making it possible to shift the paradigm in the
persistent or recurrent [3]. It seriously impairs ability to think, management of Schizophrenia. These findings are supported
perceive and behave, making it challenging to carry out daily by a comprehensive review of the literature.
Several previous studies have shown the effectiveness
The authors acknowledge the valuable support of Buch International of machine learning (ML) driven strategies in classifying
Hospital, Multan, Pakistan.
Schizophrenia. These include studies where functional mag-
979-8-3315-0486-1/24/$31.00 ©2024 IEEE netic resonance imaging (fMRI) data was used to classify
Schizophrenia using an ensemble method, which produced a scalp and provides information about the functional changes
classification accuracy of about 84% [8]. Another ensemble- associated with Schizophrenia [18] [23]. As a result of its easy
based machine learning method with neuro-anatomical MRI standardization, convenience of use for participants, sensitivity
features was used for classification with an accuracy of 83% to brain disorders, and high retest reliability, resting-state EEG
[9]. A K-Nearest-Neighbour (KNN) approach with cortical is a desirable research tool for clinical settings.
characteristics extracted from MRI data allowed for an 85%
accuracy rate in the classification of patients with Autism and A. Data Collection
Schizophrenia [10]. The works in [11] and [12] both used
Support Vector Machine (SVM) classifiers to obtain 76% and We consider publicly available resting state EEG data obtained
85% accuracies, respectively. The Multiple Kernel Learning from 14 patients of paranoid Schizophrenia at the Institute
(MKL) method used in [13] was able to achieve an even better of Psychiatry and Neurology in Warsaw, Poland [24]. The
accuracy of 86%. We summarize the performances of these patients were seven males and seven females, with ages aver-
traditional machine learning techniques in Table I below: aging 27.9 ± 3.3 and 28.3 ± 4.1 years, respectively. Patients
were included in the exclusion criteria with neurological severe
TABLE I: Performance of traditional ML Techniques conditions such as early-stage Schizophrenia, epilepsy, and
Alzheimer’s disease, in addition to other factors like pregnancy
No. Author Method Accuracy
or any general medical condition. The same institution also
1 Han [8] Ensemble Learning 84% provides a publicly available dataset of 14 healthy participants
2 Chilla [9] Ensemble Learning 83%
3 Walid Yasin [10] KNN 85% for the study who were in the same age range and gender
4 Zhu [11] SVM 76% ratio. Every subject had their EEG recorded for 15 minutes
5 Lee [12] SVM 85% while in an eyes-closed resting state. Using standard 10–20
6 Elsa Santos Febles [13] MKL 86%
EEG montage with 19 EEG channels, data was collected at
a sampling frequency of 250 Hz. Reference electrode was
For pattern recognition in EEG data, deep learning has
positioned at Fp2. The thirty-second segments are free of
emerged as a powerful technique. In recent years, it has been
artifacts (e.g., eye movements, cardiac activity, or muscle
extensively leveraged for Schizophrenia diagnosis using CNNs
contractions) used for the EEG analysis. Figure 1 illustrates
[14], Transfer Learning based techniques [15], and Long Short
the configuration in which EEG electrodes were placed on the
Term Memory-based models [16], to name a few. Particularly
heads of the subjects.
for CNN-based techniques, diagnostic performance has im-
proved over the years. Schizophrenia diagnosis using CNN
achieved an accuracy of 78% in [17]. An accuracy of 81.26%
is reported in [18] for subject-based testing while CNN models
proposed in [19], [20] and [21] report accuracies as high
as 91%, 93%, and 95%, respectively, however, they suffer
from increased requirement of trainable parameters. This is
the drawback that is addressed in this work and serves as the
main contribution of this research.
In this work, we investigate the potential of deep learning
for diagnosing Schizophrenia from resting state EEG data.
The main contribution of this work is its higher efficiency
and lower trainable-parameter requirements as compared to
existing models. The proposed CNN architecture is robust
and beats existing works on accuracy for successful EEG Fig. 1: Electrodes placement
characterization of Schizophrenia subjects. In section II, we
present the methodology used for EEG characterization along
with the proposed network architecture. Section III shows our B. Data Pre-processing
achieved results with analysis. A conclusion is presented in
section IV. The preparation steps of EEG data are critical in the overall
pipeline as they add reliability and accuracy to our findings,
II. M ATERIALS AND M ETHODS thus, enabling an effective Schizophrenia diagnosis.
EEG biomarkers can be viewed as objective inducers of 1) Signal Filtering: Data Cleaning is an essential step
cognitive impairment, a prevalent feature of schizophrenia for pre-processing as it helps to reduce noise in the data,
since they can reveal underlying neural activity [22]. The isolate specific frequency bands, enhance signal quality, and
early stages of information processing impairment in the brain ultimately facilitate diagnosis. In order to minimize distortion,
can be detected using EEG biomarkers which it can be used a Butterworth bandpass filter was used within frequency bands:
to trace the information flow in the brain. EEG measures 30–45 Hz (gamma), 4.5–7.5 Hz (theta), 2–4 Hz (delta), 8–12.5
temporal fluctuations in electrical potentials throughout the Hz (alpha) and 13–30 Hz (beta).
 2) Signal Fragmentation: By fragmenting the EEG signal that they implement convolutions, hence, reducing necessary
using fixed-length 5-second epochs with a 1-second overlap, connections which ultimately translates into lower trainable
the signals of 28 patients result in a larger dataset of 7201 EEG parameter requirement of the network. This property dis-
fragments. This aids in generalizing our work and mitigates tinguishes CNNs from the fully connected artificial neural
risks of over-fitting. The 80-20% train-test split allows for networks that are parameter-intensive. Unlike other traditional
a considerable size of 5760 EEG fragments for training and machine learning classification algorithms, CNNs have the
validation with 1441 fragments reserved for testing. ability to extract useful features from the raw data through
3) Artifact Detection and Removal: The Independent Com- various convolutional layers and pooling operations. The non-
ponent Analysis (ICA) technique is used to identify a linear linear operation of the activation functions makes such deep
transformation that reduces statistical dependence between sig- learning tools very powerful in identifying intricate patterns
nal components. This work identified statistically independent in complex datasets such as EEG.
source channels by training ICA filters on EEG data. We manifest the convolution operation within convolutional
4) Base Line Correction and Normalization: Baseline cor- layers in Equation 1 below,
rection is typically done to minimize the impact of pre- A X L
!
stimulus signals on post-stimulus signals. Normalization en-
X
Cm,k = σ Sa,f +k−1 βa,m,f + βo,m , (1)
sures the data across subjects are on a comparable scale. a=1 ℓ=1
A baseline set of 200 milliseconds was applied before the
start of every epoch up to the beginning of the epoch. The where, Cm,k refers to the convolution output for mth filter and
reference period for computing the mean signal level was this k th spatial location; σ denotes non-linear activation function;
interval in line with our baseline parameter. By performing A refers to number of input channels (A = 1 in the case of
this baseline correction in the mean mode, it was possible 1-dimensional temporal EEG data); L refers to size of the
to minimize the effect of any pre-existing signal variations convolution filter; Sa,f +k−1 refers to input feature map at
unrelated to the events of interest. Figures 2 and 3 show 19- channel a and spatial location f + k − 1; βa,m,f refers to
channel EEG profiles of healthy control and Schizophrenic the weight of the filter at f th position applied to ath input
subjects, respectively. channel for mth filter; βo,m refers to the bias term for mth
filter.
D. Proposed Model Architecture
The proposed architecture consists of a two-block
lightweight deep convolutional neural network, shown in 4:

Fig. 2: Normal EEG

Fig. 4: Proposed model architecture

• First Convolutional Block: The input layer is aligned

Fig. 3: Schizophrenic EEG with the pre-processed EEG dataset and is interfaced
with the first convolutional block of the network. The
first convolutional block is designed using a combination
C. Convolutional Neural Network Architecture of components: a convolutional layer that extracts key
These form a type of neural networks where information features and identifies temporal patterns in the EEG
is passed from one node to another in a manner similar to signals by utilizing 5 filters and a kernel size of 10; a
the feed-forward neural network. However, the connections leaky rectified linear unit (ReLU) activation function to
between neurons in hidden layers are set up in such a way avoid diminishing gradients for inactive units and thereby
 enhancing the learning efficacy; a maximum pooling
layer with a pool size of 2 so as to streamline the data
dimensionality, thus optimizing the feature representation
while crucially preserving essential information; batch
normalization to standardize layer activations which is
critical for expediting the training process and bolstering
model stability.
• Second Convolutional Block: this houses 6 filters of
kernel size 10 in its convolutional layer to further refine
the extracted features, signifying a more profound hierar-
chical processing of the EEG data. The output is flattened
to make it compatible with subsequent dense layers.
• Fully Connected Layer Block: this block contains first
fully connected layer i.e. a 32-unit dense layer, followed
by ReLU activation function, batch normalization, and
0.5 dropout layer to normalize activations and prevent
over-fitting, thus, improving ability to model to gen- Fig. 5: Accuracy and Loss Profiles
eralize. This block is interfaced with an output layer
that uses sigmoid activation function in dense layer for TABLE II: Confusion matrix of Schizophrenia classification
binary classification, showing likelihood of classification
of signals into healthy or Schizophrenic groups. Predicted Classes

III. R ESULTS AND D ISCUSSION Normal Schizophrenic

The developed CNN categorizes EEG signals into control Normal 646 16
and Schizophrenic. The training process required iterating over True
50 epochs, along with an early stopping mechanism to avoid Schizophrenic 35 744
over-fitting. The progression of training and validation losses,
along with training and validation accuracies, is shown in TABLE III: Comparison with existing works
Figure 5. A gradual increase in training accuracy is evident
No. Author Model Accuracy Parameters
from 77% to 98.40%. The gradual progression of validation
accuracy towards training accuracy in these epochs shows that 1 Oh [18] CNN 82% 118692
2 Shoeibi [20] CNN-LSTM 84% 40,028,617
there is no over-fitting or under-fitting. The strong validation 3 Barros [17] CNN 78% 1,330,873
accuracy results also add credibility to the network’s overall 4 Göker [19] CNN 91% 6,251,220
generalization. On test data, the model achieves an overall 5 Sharma [21] CNN-TCN 95% 114,297,328
6 Proposed model CNN 96% 59,972
accuracy of 96%, indicating a high level of reliability in its
predictions. The values of precision and F1-score are 0.97
and 0.96, respectively. The confusion matrix is shown in
Table II, which shows the model’s strong performance with
minimal false positives and false negatives. The ability of
the proposed model to successfully differentiate EEG signals
between normal and Schizophrenic demonstrates its potential
as a valuable tool for clinical diagnostics and investigations
into the neuro-physiological basis of Schizophrenia. Table III
provides a comparative context of proposed work in terms of
its performance with respect to the existing works in literature.
The higher accuracy with respect to the state-of-the-art models
while using fewer trainable parameters shows the robustness of
the proposed model. This difference in trainable parameters is
visually evident on a logarithmic scale in figure 6. The higher
diagnostic precision of the proposed work using comparatively
fewer trainable parameters makes the proposed model more
efficient.
Fig. 6: Comparison of trainable parameters
 IV. C ONCLUSION [14] T. Yan, G. Wang, T. Liu, G. Li, C. Wang, S. Funahashi, D. Suo, and
G. Pei, “Effects of microstate dynamic brain network disruption in
This work demonstrates superior performance of the proposed different stages of schizophrenia,” IEEE Transactions on Neural Systems
CNN architecture for Schizophrenia diagnosis as compared and Rehabilitation Engineering, vol. 31, pp. 2688–2697, 2023.
to the existing works in the literature. In addition to higher [15] D. Organisciak, H. P. Shum, E. Nwoye, and W. L. Woo, “Robin: A
robust interpretable deep network for schizophrenia diagnosis,” Expert
accuracy of proposed model, the requirement on trainable pa- Systems with Applications, vol. 201, p. 117158, 2022.
rameters of the model is also light as compared to the existing [16] M. Zhao, W. Yan, N. Luo, D. Zhi, Z. Fu, Y. Du, S. Yu, T. Jiang,
models. High specificity and sensitivity of the proposed model V. D. Calhoun, and J. Sui, “An attention-based hybrid deep learning
framework integrating brain connectivity and activity of resting-state
make it a valuable tool for early Schizophrenia diagnosis, functional MRI data,” Medical image analysis, vol. 78, p. 102413, 2022.
allowing for prompt intervention and better subsequent patient [17] C. Barros, B. Roach, J. M. Ford, A. P. Pinheiro, and C. A. Silva, “From
outcomes. This work serves as a significant step forward for sound perception to automatic detection of schizophrenia: an EEG-based
deep learning approach,” Frontiers in Psychiatry, vol. 12, p. 813460,
deep learning-driven EEG characterization with the potential 2022.
for deployment on edge devices. Extension to other neurologi- [18] S. L. Oh, J. Vicnesh, E. J. Ciaccio, R. Yuvaraj, and U. R. Acharya,
cal disorders and enhancing generalizability of the model with “Deep convolutional neural network model for automated diagnosis of
schizophrenia using EEG signals,” Applied Sciences, vol. 9, no. 14, p.
subject-specific testing is part of a future work. 2870, 2019.
[19] H. Göker, “1D-convolutional neural network approach and feature
R EFERENCES extraction methods for automatic detection of schizophrenia,” Signal,
[1] I. Chatterjee, M. Agarwal, B. Rana, N. Lakhyani, and N. Kumar, Image and Video Processing, vol. 17, no. 5, pp. 2627–2636, 2023.
“Bi-objective approach for computer-aided diagnosis of schizophrenia [20] A. Shoeibi, D. Sadeghi, P. Moridian, N. Ghassemi, J. Heras, R. Al-
patients using fMRI data,” Multimedia Tools and Applications, vol. 77, izadehsani, A. Khadem, Y. Kong, S. Nahavandi, Y.-D. Zhang et al., “Au-
pp. 26 991–27 015, 2018. tomatic diagnosis of schizophrenia in EEG signals using CNN-LSTM
[2] T. K, P. S, and A. Kumar, “Detecting schizophrenia from healthy models,” Frontiers in neuroinformatics, vol. 15, p. 777977, 2021.
control using machine learning and EEG signals,” in 2024 International [21] G. Sharma, A. M. Joshi, D. Yadav, and S. P. Mohanty, “A smart health-
Conference on Expert Clouds and Applications (ICOECA), 2024, pp. care framework for accurate detection of schizophrenia using multi-
434–442. channel EEG,” IEEE Transactions on Instrumentation and Measurement,
[3] A. A. Loch, “Schizophrenia, not a psychotic disorder: Bleuler revisited,” 2023.
Frontiers in psychiatry, vol. 10, p. 328, 2019. [22] S. K. Satapathy, P. Vyas, H. Patadia, A. Talati, and J. Patel, “Comparative
[4] S. N. Mosolov and P. A. Yaltonskaya, “Primary and secondary negative analysis of machine learning and hybrid deep learning algorithm for
symptoms in schizophrenia,” Frontiers in psychiatry, vol. 12, p. 766692, schizophrenia detection using EEG signals,” in 2024 IEEE 9th Inter-
2022. national Conference for Convergence in Technology (I2CT), 2024, pp.
[5] R. Blackmore, J. A. Boyle, M. Fazel, S. Ranasinha, K. M. Gray, 1–7.
G. Fitzgerald, M. Misso, and M. Gibson-Helm, “The prevalence of [23] N. Nattudurai, “Classifying schizophrenia disorder through EEG signal
mental illness in refugees and asylum seekers: A systematic review and analysis and machine learning,” in 2023 IEEE MIT Undergraduate
meta-analysis,” PLoS medicine, vol. 17, no. 9, 2020. Research Technology Conference (URTC), 2023, pp. 1–4.
[6] M. Dong, L. Lu, L. Zhang, Y.-S. Zhang, C. H. Ng, G. S. Ungvari, G. Li, [24] E. Olejarczyk and W. Jernajczyk, “EEG in schizophrenia,” 2017.
X. Meng, G. Wang, and Y.-T. Xiang, “Quality of life in schizophrenia: [Online]. Available: [Link]
a meta-analysis of comparative studies,” Psychiatric Quarterly, vol. 90,
pp. 519–532, 2019.
[7] L. Zhang, “EEG signals classification using machine learning for the
identification and diagnosis of schizophrenia,” in 2019 41st Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), 2019, pp. 4521–4524.
[8] S. Han, Y. Wang, W. Liao, X. Duan, J. Guo, Y. Yu, L. Ye, J. Li, X. Chen,
and H. Chen, “The distinguishing intrinsic brain circuitry in treatment-
naı̈ve first-episode schizophrenia: ensemble learning classification,” Neu-
rocomputing, vol. 365, pp. 44–53, 2019.
[9] G. S. Chilla, L. Y. Yeow, Q. H. Chew, K. Sim, and K. B. Prakash,
“Machine learning classification of schizophrenia patients and healthy
controls using diverse neuroanatomical markers and ensemble methods,”
Scientific reports, vol. 12, no. 1, p. 2755, 2022.
[10] W. Yassin, H. Nakatani, Y. Zhu, M. Kojima, K. Owada, H. Kuwabara,
W. Gonoi, Y. Aoki, H. Takao, T. Natsubori et al., “Machine-learning
classification using neuroimaging data in schizophrenia, autism, ultra-
high risk and first-episode psychosis,” Translational psychiatry, vol. 10,
no. 1, p. 278, 2020.
[11] Y. Zhu, H. Nakatani, W. Yassin, N. Maikusa, N. Okada, A. Kunimatsu,
O. Abe, H. Kuwabara, H. Yamasue, K. Kasai et al., “Application of
a machine learning algorithm for structural brain images in chronic
schizophrenia to earlier clinical stages of psychosis and autism spectrum
disorder: a multiprotocol imaging dataset study,” Schizophrenia Bulletin,
vol. 48, no. 3, pp. 563–574, 2022.
[12] L.-H. Lee, C.-H. Chen, W.-C. Chang, P.-L. Lee, K.-K. Shyu, M.-
H. Chen, J.-W. Hsu, Y.-M. Bai, T.-P. Su, and P.-C. Tu, “Evaluating
the performance of machine learning models for automatic diagnosis
of patients with schizophrenia based on a single site dataset of 440
participants,” European Psychiatry, vol. 65, no. 1, p. e1, 2022.
[13] E. Santos Febles, M. Ontivero Ortega, M. Valdés Sosa, and H. Sahli,
“Machine learning techniques for the diagnosis of schizophrenia based
on event-related potentials,” Frontiers in Neuroinformatics, vol. 16, p.
893788, 2022.