MambaStress: A Lightweight Multimodal State-Space Model for Real-Time Stress Detection on Wearable Edge Devices

Abstract

Continuous stress detection from wearable physiological signals is vital for proactive mental health management. Existing approaches suffer from single-modality dependence, lenient subject-mixed evaluation, and architectures too large for edge deployment. This paper introduces MambaStress, a multimodal deep learning framework that addresses all three limitations simultaneously by fusing four physiological channels-electrocardiogram, electrodermal activity, respiration, and blood volume pulse through per-signal depthwise-separable CNN encoders followed by a Mamba-based selective state-space fusion block with linear temporal complexity. On the WESAD benchmark under strict subject-independent evaluation, MambaStress achieves 95.16% accuracy and 98.44% ROC-AUC on an 80/20 held-out split, and 91.22% ± 10.31% under full leave-one-subject-out cross validation across 15 participants. With only 61,058 parameters (0.23MB FP32), the INT8-quantized classifier head runs in 1.85 ms on an ESP32 microcontroller under a split-inference setting with 322KB RAM free, confirming practical edge suitability. A signal-level ablation study demonstrates the complementary contribution of each physiological modality to overall classification performance, providing physiological insights into multimodal autonomic stress responses