On-Device Deep Learning with Live Channel Attention for Real-Time ANS State Classification on Wearable Platforms
Anandhu P.1, Ajay B. S.2 and Arun C.3
1 Department of Artificial Intelligence and Data Science, Nehru Institute of Engineering and Technology, Coimbatore
2 Department of Artificial Intelligence and Data Science, Nehru Institute of Engineering and Technology, Coimbatore
3 Department of Artificial Intelligence and Data Science, Nehru Institute of Engineering and Technology, Coimbatore
Email: 1 anandhupadmanabhank@gmail.com, 2 ajayajayprimary@gmail.com, 3arunchandrasekaran0218@gmail.com
Abstract—Edge-deployed wearable systems for autonomic nervous system (ANS) monitoring predominantly rely on handcrafted statistical features and binary thresholds, yielding no real-time insight into which physiological channel drives each classification decision. A deep learning platform is presented that addresses these dual limitations directly on an ESP32 microcontroller without cloud dependency. Four physiological modalities are acquired concurrently: electrodermal activity via a galvanic skin response (GSR) sensor, peripheral oxygen saturation (SpO2) and heart rate via a photoplethysmographic pulse oximetry module, skin surface temperature via a precision analog temperature sensor, and three-axis linear acceleration via a microelectromechanical systems (MEMS) inertial sensor. Raw 10-second multimodal windows are classified by a quantized one-dimensional convolutional neural network with a bidirectional long short-term memory layer (1D-CNN + BiLSTM) into four clinically distinct ANS states: normal baseline, sympathetic arousal, parasympathetic suppression, and mixed dysregulation. A novel Physiological Autonomic Signature Token (PAST)—a 4-dimensional learned Channel Attention Vector (CAV) requiring only 256 additional parameters—is computed at every inference pass and displayed live on the device LCD, delivering the first inference-time, per-sensor attribution on a microcontroller-resident physiological classifier. A Physiological Coherence Score (PCS), computed as the cosine similarity between the BiLSTM hidden state slices of the two dominant PAST channels and requiring no additional parameters, provides a third confidence dimension that detects inter-channel physiological inconsistency and flags suspected sensor artifacts at inference time. Monte Carlo (MC) Dropout over T = 20 stochastic forward passes provides predictive uncertainty quantification; high-variance predictions trigger a 'LOW CONFIDENCE' alert. The system sustains approximately 47 hours of monitoring at 42 mA average current from a 2000 mAh cell, completing single-pass inference in 6.8 ms—advancing TinyML for interpretable, confidence-aware ANS monitoring.
Index Terms—Autonomic nervous system monitoring, channel attention, edge inference, explainable AI, TinyML, wearable health monitoring.
