Wearable Holter monitors and adhesive ECG patches stream gigabytes of raw electrophysiological telemetry. Detecting paroxysmal atrial fibrillation and ventricular tachycardia without draining battery power requires ultra-compact 1D convolutional neural networks deployed on ARM Cortex-M33 silicon. This study details latency, precision, and clinical validation under ANSI/AAMI EC57.
By converting FP32 model weights to INT8 fixed-point representation with post-training quantization (PTQ), RAM utilization is reduced from 420 KB to 58 KB with a sensitivity degradation of less than 0.12% across MIT-BIH Arrhythmia benchmark datasets.
Continuous BLE radio transmission depletes coin-cell batteries in under 48 hours. The on-device SaMD model holds radio silence during normal sinus rhythm, activating high-bandwidth raw telemetry streaming only when morphological anomalies or rhythm irregularities cross confidence thresholds.