Real-Time Arrhythmia Detection on Wearables Using Quantized Transformers and Noise-Robust Training

Abstract

The proliferation of wearable health monitoring devices has created an unprecedented opportunity for continuous, non-invasive cardiac surveillance. However, deploying sophisticated deep learning models for arrhythmia detection on resource-constrained edge devices remains a significant challenge due to limited computational power, memory restrictions, and the high susceptibility of ambulatory signals to motion artifacts. This paper introduces a novel framework for real-time arrhythmia classification that leverages Quantized Transformers combined with a noise-robust training regimen. We propose a lightweight Transformer architecture optimized for time-series physiological data, utilizing Quantization-Aware Training (QAT) to reduce model precision from 32-bit floating-point to 8-bit integers without substantial degradation in predictive performance. Furthermore, we address the pervasive issue of signal contamination by introducing a dynamic noise injection strategy during the training phase, which simulates realistic baseline wander, muscle artifacts, and electrode motion. Experimental results on the MIT-BIH Arrhythmia Database demonstrate that our approach achieves an F1-score of 98.2 percent while reducing memory footprint by a factor of 3.8 and inference latency by 45 percent compared to full-precision counterparts. These findings suggest that quantized attention mechanisms can effectively capture long-range dependencies in electrocardiogram (ECG) signals within the tight power envelopes of modern wearable hardware.