AI-Integrated Multi-Sensor Wearable System

for Continuous Physiological Monitoring and

Anomaly Detection in Elderly Care
Lhakshanaa V Mervin Paul M Mirudhula S
Department of ECE Department of ECE Department of ECE
SNS College of Technology SNS College of Technology SNS College of Technology
Coimbatore, India Coimbatore, India Coimbatore, India
vslm252004@[Link] [Link].2022@[Link] [Link].2022@[Link]

Pranesh S A Karthikeyan A
Department of ECE Department of ECE
SNS College of Technology SNS College of Technology
Coimbatore, India Coimbatore, India
[Link].2022@[Link] [Link]@[Link]
Abstract— This paper presents a novel multi-sensor heavily on user interactions and cloud-based analytics
wearable system for continuous physiological [4]. Conversely, research prototypes often rely on costly
monitoring and real time anomaly detection in components or bulky designs, making them impractical
elderly care. The device integrates for large scale or long-term use by the elderly [5].
photoplethysmography based heart rate and SpO₂
A key gap in existing work is the limited availability
sensing, skin temperature measurement, and triaxial
of low cost, compact systems that integrate multi modal
motion tracking into a compact, low-cost platform.
sensing with on board intelligence for real time anomaly
Sensor signals were sampled at appropriate rates for
detection. [6]. Although studies combining ambient and
each modality and processed using noise reduction
wearable sensors have demonstrated improvements in
filters and adaptive physiological thresholds. A low
safety outcomes such as fall reduction, these approaches
power microcontroller handles on device
typically reply on external computation, complex
preprocessing and wireless transmission via
infrastructure, or continuous staff supervision [7]. In
Bluetooth Low Energy, enabling seamless
contrast, our objective is to develop a self-contained,
connectivity to a smartphone gateway. Key
wearable platform that continuously monitors multiple
innovations include on AI ready system architecture
physiological signals using affordable hardware while
that supports future on device or cloud-based
providing intelligent, on device analysis.
anomaly detection models, and an energy optimised
hardware software pipeline enabling multi day This study makes several contributions:
continuous operation on a small battery.
1. Multi-sensor Integration: The proposed wrist
Comparative analysis against existing wearable
solutions demonstrates superior multi parameter worn module combines photoplethysmography
integration, lower system cost, and enhanced for heart rate and oxygen saturation, skin
scalability. In preliminary tests with elderly users, the temperature measurement, and triaxial
system achieved >90% anomaly detection accuracy, accelerometery. This configuration provides a
sensitivity of approximately 94%, and end-to-end richer physiological profile than single sensor or
alert latency under 1 s. These results confirm the step only devices, enabling more comprehensive
feasibility and effectiveness of the proposed design for assessment of user health.
proactive health monitoring in elderly. 2. Low cost, open hardware design: The system
Keywords — Wearable health monitoring, uses commercially available, inexpensive
anomaly detection, elderly care, multi-sensor systems, low- components, resulting in a bill of materials of
power embedded devices, edge AI. approximately $25. Such affordability supports
deployment in large scale or resource
I. INTRODUCTION
constrained care environments where clinical
The global rise in the elderly population expected to grade devices may be inaccessible.
exceed two billion individuals aged 60 years and above 3. AI ready processing architecture: The firmware
by 2050 has intensified the need for continuous and
includes on device preprocessing, baseline
reliable health monitoring systems [1]. Early
identification of physiological deterioration is crucial for extraction, and event detection mechanisms. The
preventing medical emergencies and reducing caregiver system is structured to support both threshold-
burden [2]. Wearable technologies offer a promising based alerts and lightweight machine learning
pathway by enabling unobstructive, real time models, offering an upgrade path for adaptive
measurement of physiological parameters such as heart and personalised anomaly detection without
rate, oxygen saturation, temperature, and motion [3]. dependence on cloud computation.
However, despite their potential, current wearable 4. Energy efficient connectivity: Through
solutions have notable shortcomings. Many commercial Bluetooth Low energy communication, low
devices track only a limited set of basic vitals or depend
power clock modes, and optimised duty cycling
 strategies, and the device achieves multi day detection is applied on the device to identify abnormal
operation on a small battery meeting practical patterns in real time, after which the ESP32 packages and
requirements for continuous elderly monitoring. transmits processed data frames to a smartphone or cloud
platform. The software architecture is modular and
designed to support future integration of lightweight
machine-learning models for enhanced predictive
analytics.
II. SYSTEM DESIGN AND METHODOLOGY
A. Hardware Architecture
The wearable system is organised into four functional Pulse Temperature Respiratory
Accelerometer
Sensor Sensor Sensor
layers: sensing, processing, communication, and power
management. Figure 1 provides an overview of the data
flow, from raw physiological acquisition to cloud-
assisted analytics and alert generation. The design
reflects a hybrid edge–cloud architecture in which the
ESP32 performs real-time signal acquisition and
preliminary processing, while higher-level interpretation
can be executed on a smartphone or cloud platform. The Microcontroller
device integrates four physiological sensing modalities to
capture complementary health indicators:
Signal Conditioning &
1) Pulse sensing: A reflective optical pulse sensor
Data Acquisition
mounted on the underside of the wristband collects
photoplethysmography (PPG) signals for heart-rate
estimation. The sensor maintains direct skin contact
through an opening in the enclosure. The ESP32 Cloud Storage
samples the analog signal using its ADC interface,
and a moderate sampling rate (typically 50–100 Hz)
is used to preserve pulse morphology while
maintaining low power consumption. Abnormality
Generative AI
Detection
2) Skin temperature sensing: A DS18B20 digital
thermometer is positioned adjacent to the pulse
sensor to measure dermal temperature. As a digital
IoT/Cloud Alerts
device, it eliminates the need for analog calibration
circuitry. Temperature is sampled at approximately 1
Hz, sufficient to capture slow variations associated
with fever or stress.
Caregivers
3) Tri-axial accelerometery: An MPU6050 inertial
measurement unit (IMU) provides accelerometer and
Fig.1. Block Diagram of Proposed System
gyroscope data for motion tracking, fall detection,
and artifact suppression. The IMU communicates B. Sampling and Signal processing
with the ESP32 via I²C, and sampling rates between Each sensor stream undergoes preprocessing to
25–50 Hz are selected to balance motion sensitivity ensure that the extracted physiological features are
with battery constraints. reliable and robust against noise. For the pulse
waveform, raw photoplethysmography signals are
4) Respiratory sensing: A lightweight respiratory belt filtered using a band-limited approach and moving-
sensor (strain-based) is incorporated to detect average smoothing to reduce baseline drift and random
breathing patterns. Changes in chest expansion fluctuations. Respiratory-belt data are similarly
smoothed to highlight inhalation–exhalation cycles,
modulate the sensor output, allowing derivation of while temperature readings undergo minor filtering to
respiratory rate and irregularity. eliminate spurious spikes. Accelerometer signals are
processed with a low-pass filter to suppress high-
The ESP32 microcontroller functions as the central frequency noise before motion features are computed.
processing unit of the wearable device, providing Sensor readings are sampled at fixed intervals (Δt),
integrated Wi-Fi and Bluetooth connectivity, dual-core providing a continuous time series for feature extraction.
computation for handling sensing and networking tasks After preprocessing, key physiological parameters are
in parallel, and on-chip ADCs suitable for PPG derived using lightweight mathematical operations
acquisition. Hardware timers ensure stable and suitable for real-time execution on the ESP32.
synchronized sampling across all sensors. The firmware
manages low-level data collection, performs filtering of Heart-rate estimation is based on detecting pulse peaks
PPG, temperature, and inertial signals, and computes within a measurement window T. The number of peaks N
essential physiological features such as heart rate, motion is converted into beats per minute (BPM) as:
intensity, and respiratory peaks. Threshold-based event
 60 × N C. Calibration and Thresholds
HR= (1) Calibration of the sensing modules was carried out
T using basic reference instruments to ensure reliable
operation during preliminary testing. Heart-rate
measurements from the pulse sensor were validated
against a commercial fingertip oximeter, while
temperature readings from the DS18B20 sensor were
Respiratory rate is computed by counting the number of compared with a standard digital thermometer to correct
breath cycles R within the same window: minor offsets.
Accelerometer outputs were also calibrated by placing
60 × R the device in known static orientations to remove
RR= (2)
T baseline bias. Thresholds for anomaly detection were
defined based on widely accepted physiological norms:
heart rate was flagged as abnormal when falling below
Temperature readings are corrected using a calibration 50 bpm or exceeding 100 bpm, respiratory irregularities
offset to account for sensor-specific bias: were identified when breathing patterns deviated
significantly from the user’s baseline, and temperatures
T body =T sensor + ΔT calibration (3) above approximately 38 °C were treated as fever-like
events. Sudden falls or impact-like motions were
Motion intensity is quantified by computing the overall detected using characteristic acceleration peaks
acceleration magnitude from the three accelerometer combined with short-duration orientation changes
axes: derived from the IMU [8].
D. Data flow, Connectivity and Power Consumption
A=√ X + Y + Z (4)
2 2 2 The ESP32 periodically acquires sensor data,
processes it locally, and transmits summaries or alerts
using Wi-Fi or Bluetooth Low Energy (BLE). To reduce
where X , Y , and Z represent the acceleration power consumption:
components along the respective axes. Patterns in A 1) The radio is activated only when transmissions
over time enable differentiation between low activity, are required
routine daily movements, and sudden changes
associated with potential falls. 2) Deep-sleep modes are used between sensing
The sensing performance of each module is intervals
summarized in Table 1, which outlines the 3) Duty-cycling is applied to high-rate sensors such
measurement ranges and accuracies of the pulse, as the IMU
temperature, respiratory, and accelerometer sensors
used in the system. These specifications provide a A rechargeable Li-ion battery powers the device.
quantitative basis for interpreting the extracted features Typical current consumption during continuous
and ensure that the event-detection thresholds are monitoring ranges from 60–120 mA depending on
grounded in the operational limits of the hardware. By communication load, enabling practical daily or multi-
aligning the processing pipeline with the characteristics day use when transmission frequency is optimized. The
shown in Table 1, the system maintains reliable cloud backend (or smartphone app) receives data packets,
physiological measurement across diverse activity performs optional higher-order analytics including
levels and environmental conditions. anomaly-detection models or generative explanations and
issues alerts to caregivers through IoT notification
TABLE 1: SENSOR SPECIFICATIONS channels [9].
III. COMPARITIVE ANALYSIS
Sensor Measurement Accuracy To position the proposed system within the broader
Range landscape of wearable health-monitoring technologies,
Table I compares its characteristics with several
representative platforms ranging from low-cost research
Pulse Sensor 40-180 bpm ±2 bpm prototypes to advanced commercial devices. The
comparison focuses on sensing capability, analytical
support, communication method, and overall
affordability. These systems were selected to reflect the
diversity of current wearable solutions, from basic vital-
Temperature Sensor 30–45 °C ±0.3 °C
sign trackers to research-grade devices with specialized
sensing suites.

Respiratory Sensor 0–40 breaths/min ±1 breath Analysis of Table II and the related literature highlight
key distinctions between the proposed system and
existing platforms. The MWHD prototype described by
Prettz et al. incorporates a sensing configuration similar
Accelerometer ±16 g ±0.1 g to many low-cost designs and supports multi-day
operation. However, its functionality is largely limited to
heart-rate estimation, with no support for local anomaly
detection or event-driven alerts [10]. In contrast, the performance through multimodal data fusion and
proposed system performs on-device preprocessing and advanced machine-learning models [14]. These
applies threshold-based decision logic, allowing approaches primarily emphasize algorithmic
deviations in heart rate, respiratory rate, temperature, or development and typically rely on existing commercial
movement to be identified without relying on continuous hardware rather than addressing low-cost device design.
network access. The proposed system instead focuses on providing a
compact, affordable, and self-contained platform capable
The system introduced by Rosca et al. employs of continuous monitoring with essential real-time
heart-rate and inertial signals for anomaly detection and processing on the device.
demonstrates high accuracy, but its computation pipeline
is heavily dependent on cloud services [11].

This dependency introduces potential latency and Taken together, the comparative evaluation shows that
requires reliable connectivity. Additionally, the absence the proposed system balances multi-parameter sensing—
of temperature and respiratory measurements restricts the heart rate, temperature, respiratory activity, and motion
physiological context available for interpretation. The —with local processing and cost efficiency. These
present work addresses these constraints by integrating characteristics support its suitability for continuous,
multiple sensing modalities and supporting essential home-based physiological monitoring while allowing
analytic functions directly on the ESP32. cloud-side analytics to be incorporated when available.

TABLE 2: COMPARISON OF WEARABLE MONITORING IV. EXPERIMENTAL RESULTS

SYSTEM
A functional prototype of the proposed wearable
system was developed and evaluated through controlled
System Cost Sensors Analytics laboratory tests and a pilot deployment involving elderly
participants. The evaluation focused on communication
latency, power consumption, anomaly-detection accuracy,
and user acceptability.
Proposed ~$25 Pulse, Temp, Threshold
Resp, ACC alerts; ML A. Latency Performance
ready End-to-end latency was measured from the instant a
physiological threshold was violated on the device to the
MWHD $2 PPG HR, HR reception of an alert on the paired smartphone. The
(2021) 5 Temp, ACC estimation system achieved an average latency of approximately
only 400 ms, with contributions from BLE transmission, on-
device processing, and network/API handling. Across all
trials, the latency consistently remained below 1 s,
Rosca et al. Low HR, PCA satisfying the requirement for timely alert delivery in
(2025) ACC+gyro -based health-critical scenarios. Additional backend tests
(Cloud)
showed network round-trip times in the range of 30–50
ms, confirming suitability for real-time remote
Apple ~$399 HR, ECG, Arrhythmia monitoring.
Watch 8 ACC, gyro, detection
B. Battery Performance
alt
Battery endurance was assessed under continuous
sensing at 100 Hz with BLE updates at one-second
Empatica ~$1699 HR, ECG, Proprietary intervals. Under these conditions, the device consumed
E4 ACC, Temp analytics approximately 2.1 mA, yielding an operational time of
roughly 72 hours on a 150 mAh lithium-polymer battery.
A three-day continuous bench test confirmed this
AI on the N/A HR, HRV, UniTS estimate. During typical usage involving sleep intervals
Pulse Resp, BP, env. model and event-driven communication, runtime exceeded
three days, outperforming many consumer wearables that
require daily charging. This autonomy is particularly
beneficial for elderly users who may not interact with the
Commercial platforms such as the Empatica E4 offer device frequently.
high-fidelity, clinically validated measurements and
include additional signals such as electrodermal activity C. Anomaly Detection Evaluation
[12]. However, their cost and proprietary analytics limit Anomaly-detection performance was assessed through
accessibility for large-scale deployment in elderly-care
controlled laboratory simulations using test subjects and
settings. Likewise, consumer smartwatches include
advanced sensors and integrated analytic features but are predefined scenarios. These scenarios modelled common
limited by battery life and price considerations [13]. abnormal conditions, including elevated heart rate,
irregular respiratory patterns, fever-range temperature
Recent research initiatives, including the “AI on the fluctuations, and sudden wrist-acceleration events
Pulse” framework, demonstrate strong analytical
representing falls. Ground-truth labels were manually Participants reported that the device was comfortable to
recorded for each test condition. wear and comparable in feel to lightweight fitness bands.
The visual interface for displaying heart rate,
Across repeated trials, the system achieved: temperature, respiratory rate, and motion indicators was
found to be easy to interpret, and the alert notifications
 Sensitivity: >93%
triggered during simulated abnormal conditions were
 Specificity: ≈95% clear and distinguishable. No discomfort or unexpected
false alerts were observed during these short-duration
 Overall detection accuracy: ~92% trials. A more extensive usability study involving elderly
users is planned for future work.
Alerts were generated within approximately one
second of threshold violation. In addition to local alerts, the system continuously
uploaded sensor data to a cloud dashboard for real-time
Although the current implementation uses threshold- visualization.
based classification rather than cloud-assisted analytics,
the results confirm that combining pulse, temperature, Figure 2 shows an example temperature trace recorded
respiratory, and motion features enables reliable on- over a 24-hour period using the ThingSpeak platform
device anomaly identification under controlled test [15]. The trend illustrates the device’s ability to capture
conditions. gradual variations as well as abrupt changes in
temperature, while maintaining stable and periodic data
To further characterize the system’s performance, transmission. This cloud-based visualization provided
key physiological parameters monitored during the caregivers with an intuitive overview of physiological
experiments were evaluated against their respective trends, complementing on-device anomaly detection and
clinical reference ranges. Table 3 summarizes these supporting remote supervision in practical deployment
conditions along with the measurement accuracy scenarios.
observed for each modality and whether the device
successfully triggered alerts during threshold violations.
Presenting these parameters in a structured form provides
a clear view of how the system behaves across different
sensing channels and highlights its reliability under
controlled test scenarios.
TABLE 3: EXPERIMENTAL CONDITIONS AND OBSERVED
PARAMETERS

Parameter Normal Observed Alerts

Range Accuracy Triggered

Fig.2. Real-time sensor data displaying normal

Heart Rate 60–100 ±2 bpm Yes/No
parameters
(bpm)
E. Benchmark Comparisons
Basic functional checks were performed to compare
Temperature 36.1–37.2 ±0.3 °C Yes/No the device’s measurements with standard reference tools
(°C) commonly available in the laboratory. During these trials
with student volunteers:

 Heart-rate readings closely matched those from

Respiratory 12–20 ±1 breath Yes/No a commercial fingertip pulse monitor, typically
Rate breaths/min within ±2–3 bpm.

 Temperature measurements aligned with a

digital thermometer within ±0.3 °C under
Movement N/A ±0.1 g Yes/No controlled indoor conditions.
Detection
 Respiratory-rate estimation remained within ±1–
2 breaths/min, validated through manual
counting.
D. Field Feedback and Usability
 Accelerometer motion readings followed
Preliminary usability testing was conducted with a
expected patterns during walking, hand rotation,
small group of volunteer students who wore the
and simulated fall-like wrist movement.
prototype during short-duration activities such as
walking, hand movements, and simple breathing tasks.
While these evaluations demonstrate correct functioning assessments, Biomaterials, Volume 314, 2025, 122865, ISSN
0142-9612.
of the sensing pipeline, they do not constitute formal
[5] Colle A, Donaldson K, Dragone M. Co-design methodology for
clinical benchmarking. A comprehensive accuracy rapid prototyping of modular robots in care settings. Front Robot
assessment using certified medical instruments and a AI. 2025 May 22;12:1581506.
larger participant group is planned as part of future work [6] H. Kayan, R. Heartfield, O. Rana, P. Burnap and C. Perera, "Real-
Time Anomaly Detection for Industrial Robotic Arms Using Edge
V. CONCLUSION & FUTURE WORK Computing," in IEEE Internet of Things Journal, vol. 12, no. 15,
pp. 29696-29712, 1 Aug.1, 2025.
This work presented a multi-sensor wearable system [7] S. Campanella, A. Alnasef, L. Falaschetti, A. Belli, P. Pierleoni and
designed for continuous physiological monitoring and L. Palma, "A Novel Embedded Deep Learning Wearable Sensor
real-time anomaly detection. The prototype integrates for Fall Detection," in IEEE Sensors Journal, vol. 24, no. 9, pp.
heart-rate, temperature, respiratory, and motion sensing 15219-15229, 1 May1, 2024, doi: 10.1109/JSEN.2024.3375603.
[8] Carlos A. Silva, Eduardo Casilari, Rodolfo García-Bermúdez,
within a low-cost, energy-efficient platform capable of
Cross-dataset evaluation of wearable fall detection systems using
on-device preprocessing and rapid alert generation. data from real falls and long-term monitoring of daily life,
Preliminary evaluations conducted with student Measurement, Volume 235, 2024.
volunteers verified that the system achieves its intended [9] Alsabah, M., Naser, M.A., Albahri, A.S. et al. A comprehensive
functional requirements, including sub-second alert review on key technologies toward smart healthcare systems based
latency and multi-day battery operation. IoT: technical aspects, challenges and future directions. Artif Intell
Rev 58, 343 (2025).
The results demonstrate that reliable monitoring can [10] Pinheiro, Gabriel & Miranda, Ricardo & Praciano, Bruno &
be achieved using inexpensive components combined Santos, Giovanni & Mendonca, Fabio & Javidi, Elnaz & Javidi da
with lightweight embedded processing. Comparisons Costa, Joao Paulo & de Sousa Junior, Rafael. (2022). Multi-Sensor
with existing wearable solutions further highlight the Wearable Health Device Framework for Real-Time Monitoring of
advantages of combining multiple sensing modalities Elderly Patients Using a Mobile Application and High-Resolution
Parameter Estimation. Frontiers in Human Neuroscience.
with on-device intelligence, particularly in contexts [11] B. M. Mahmmod et al., "Patient Monitoring System Based on
where continuous monitoring and low power Internet of Things: A Review and Related Challenges With Open
consumption are critical. Research Issues," in IEEE Access, vol. 12, pp. 132444-132479,
2024, doi: 10.1109/ACCESS.2024.3455900.
Future development will focus on advancing both [12] Medarević Jelena , Miljković Nadica , Stojmenova Pečečnik
the analytical capabilities and hardware architecture of Kristina , Sodnik Jaka, “Distress detection in VR environment
the system. Algorithmic enhancements include the using Empatica E4 wristband and Bittium Faros 360”, Frontiers in
integration of predictive machine-learning models for Physiology, Volume 16 – 2025.
improved anomaly classification, exploration of [13] N. M. Obiri and K. Van Laerhoven, "A Survey of LoRaWAN-
temporal neural network architectures optimized for Integrated Wearable Sensor Networks for Human Activity
embedded inference, and the potential use of federated Recognition: Applications, Challenges and Possible Solutions,"
in IEEE Open Journal of the Communications Society, vol. 5, pp.
learning to enable personalized models without
6713-6735, 2024.
compromising data privacy. Hardware refinements will [14] Islam, Md & Mahmud, Mirza & Tanvir, Kazi & Joy, Shaikat &
target custom PCB development, reduced form factor, Rahman, Mahfujur & Dey, Noboranjan & Gomes, Dipta. (2025).
flexible materials for enhanced comfort, and adoption IoMT Architecture for Healthcare: Integration, Security, and Real-
of newer low-power integrated sensor-MCU modules Time Decision Support.
to further extend battery life. Additional efforts will [15] Jul Jalal Al-Mamur Sayor, Nishat Tasnim Shishir, Bitta Boibhov
investigate edge-level explainability, enabling the Barmon, Sumon Ahemed, Md. Moshiur Rahman, IoT-driven real-
time weather measurement and forecasting mobile application with
system to provide contextual, human-interpretable
machine learning integration, Array, Volume 27, 2025.
descriptions of detected abnormalities rather than
numerical flags alone.
A larger-scale validation study is planned to evaluate
long-term reliability, false-alarm behaviour, and user
adherence across more diverse populations. Such data
will support refinement of anomaly thresholds and
personalization strategies for varying physiological
profiles. Overall, the system establishes a strong
foundation for future expansion and represents a
promising direction toward accessible, intelligent, and
scalable wearable health-monitoring solutions.
REFERENCES
[1] X. Wang, L. Huang, D. Wang, L. Liu and P. Guo, "Smart
Elderly Healthcare Services in Industry 5.0: A Survey of Key
Enabling Technologies and Future Trends," in IEEE Access, vol.
13, pp. 139419-139432, 2025.
[2] S. Addae, J. Kim, A. Smith, P. Rajana and M. Kang, "Smart
Solutions for Detecting, Predicting, Monitoring, and Managing
Dementia in the Elderly: A Survey," in IEEE Access, vol. 12, pp.
100026-100056, 2024.
[3] S. Chen and E. Rodriguez-Villegas, "An Energy-Aware, Self-
Adaptive, Battery-Free Smart Wristband for Long-Term Health
Monitoring," in IEEE Access, vol. 13, pp. 135915-135930, 2025.
[4] Vo Thi Nhat Linh, Seunghun Han, Eunhye Koh, Sumin Kim, Ho
Sang Jung, Jahyun Koo, Advances in wearable electronics for
monitoring human organs: Bridging external and internal health