StressNet: Neuromorphic multimodal stress recognition from wearable physiological signals

Continuous, real-time monitoring of psychological stress is becoming increasingly important in digital health and remote patient-monitoring systems. However, deploying always-on stress-monitoring models on consumer wearable devices remains challenging because of their limited computational resources and microjoule-scale energy budgets. This paper presents StressNet (Multimodal Spiking Stress Network), a spike-driven neuromorphic framework for real-time, subject-independent stress monitoring from multimodal physiological signals acquired using a consumer wrist-worn device.

StressNet is designed for efficient on-device inference on edge-class wearable hardware. Each physiological modality is first converted into spike trains using a learnable adaptive-threshold leaky-integrate-and-fire encoder. Reliability-gated sparse spike-feature cross-modal attention is then used to selectively integrate complementary information across modalities, while a recurrent neuromorphic memory captures temporal dependencies across time.

The network is trained end-to-end using surrogate-gradient backpropagation through time and evaluated under leave-one-subject-out (LOSO) cross-validation on the three-class WESAD benchmark using only the six wrist-based physiological channels. Across five random seeds, StressNet achieves a mean accuracy of 97.50% with a standard deviation of 0.88 percentage points and a macro-F1 score of 0.974. These results were obtained using 60-second windows with a 7-second stride.

A mean expected calibration error (ECE) of 0.023 indicates good calibration under the evaluated protocol. StressNet requires an estimated 32.5,\(\mu\)J of energy and 7.95,ms of model-side inference time per window. With preprocessing included, these estimates increase to approximately 33.5,\(\mu\)J and 8.21,ms.

These values are analytical estimates rather than direct hardware measurements. Overall, StressNet demonstrates promising subject-independent stress-recognition performance and estimated computational efficiency for wearable deployment, while broader calibration studies and real-world hardware validation remain important directions for future work. The authors acknowledge the use of AI-based language-assistance tools, including the latest versions of Grammarly and Writefull for Overleaf, solely to refine the language and grammar of the manuscript (spelling, wording, and readability).

The authors declare that this research received no external funding. Department of Electrical and Computer Engineering, George Mason University, Fairfax, 22030, VA, USA College of Engineering and IT, University of Dubai, Dubai, UAE Software Engineering Department, Faculty of Informatics, Kaunas University of Technology, Kaunas, 51423, Lithuania Correspondence to Muhammad Zaheer Sajid. The authors declare no competing interests.