Into the Wild: Reliable Physiological Sensing with on-device Autoencoder-based Anomaly Detection

Kevin Penner, Felix Wittenfeld, Marc Hesse, Michael Thies

Confirmation: I have read and agree with the IEEE BSN 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: wireless body sensor, wearable, autoencoder, ecg anomaly, signal quality index (SQI), ultra-low-power microcontroller, TinyML, on-device inference

Abstract: Pushing physiological sensing into the wild: this work equips our ultra-low-power physiological BI-Vital sensor with an $\textit{int8}$-quantized autoencoder that produces a real-time signal-quality index (SQI) for electrocardiograms (ECG) beside standard heart rate computations. The supervised model, trained on PhysioNet CinC2017 and three newly annotated BI-Vital single-lead ECG datasets, combines a compact encoder-decoder with a lightweight classifier head. On the STM32L476JE microcontroller from BI-Vital, the model uses 31 kB of RAM and 85 kB of flash memory, completes inference in 195 ms and still leaves room for parallel sensor services. Across eight train/test splits the approach generalizes well, reaching a F1 score of 0.988 while maintaining $2 \cdot 10^{-3}$ performance loss after quantization. By converting raw continuous ECG data into periodic heart rate and SQI recordings, storage space on the device is reduced by 99.6 \%, demonstrating a practical way for reliable, data-efficient monitoring outside the laboratory.

Track: 3. Signal processing, machine learning, deep learning, and decision-support algorithms for digital and computational health

NominateReviewer: Marc Hesse: [e-mail]

Submission Number: 69