Hardware-Aware Edge Deep Learning for Multimodal Biomedical Monitoring
Multi-task CNN–SE–Transformer running entirely on an ESP32-S3 for real-time ECG arrhythmia, stress, and activity classification — with on-device adaptive learning and motion-aware alerting. No cloud required.
- 112,552-parameter multi-task model deployed uncompressed on ESP32-S3, at 0.54 W for the complete sensing system
- 715× reduction in activation storage via head-only backpropagation (4,984 trainable vs. 107,568 frozen parameters)
- Patent-pending EdgeGuard on-device adaptation framework
- Leave-one-subject-out cross-validation across 65 subjects (arrhythmia 92.4%)