Deep Learning on the Edge: Highly Compressed Multi-Scale CNNs for Wearable Stress Detection via Knowledge Distillation

This work presents a lightweight deep learning pipeline for wearable stress detection designed with edge deployment constraints in mind. The approach uses a Multi-Scale 1D-CNN Teacher with parallel convolutional branches to capture physiological patterns occurring at different temporal scales across ECG, EDA, respiration, temperature, EMG, and acceleration signals. A response-based Knowledge Distillation framework is then used to transfer the Teacher's predictive knowledge to an ultra-lightweight MicroCNN Student containing only 5,352 trainable parameters. On the WESAD dataset, evaluated using Leave-One-Subject-Out cross-validation, the distilled MicroCNN achieves 99.8% accuracy and an F1-score of 0.996, compared with an F1-score of 0.988 for the 266K-parameter Teacher. This represents approximately 50× parameter compression while maintaining and improving predictive performance. The work also examines how compression affects the model's decision strategy using SHAP-based explainability. The analysis shows that the Student develops a different feature-attribution pattern from the Teacher, shifting from an ECG-dominant strategy toward greater reliance on temperature and EDA features while maintaining strong classification performance. Finally, the distilled model is exported to ONNX and benchmarked on a Raspberry Pi 4, achieving a mean inference latency of 2.38 ms with a 24.3 KB model size. This preprint documents the methodology, experimental evaluation, explainability analysis, and edge deployment results of the project.