Multihead Attention-Based Denoising for Enhanced ECG Signals in Wearable Biosensors

Wearable biosensors for ECG monitoring face significant challenges in achieving clinical-grade signal quality due to various noise artifacts, including motion interference, baseline wander, and powerline noise. This study introduces a Multi-Head Attention Denoising (MHAD) method, leveraging its ability to capture intricate signal dependencies across time steps and frequencies. Unlike conventional approaches, MHAD processes diverse aspects of noisy ECG signals in parallel through multiple attention heads, enabling precise noise suppression while preserving ECG critical features. Comprehensive evaluations demonstrate substantial improvements in key metrics, including a Signal-To-Noise Ratio (SNR) of 24.46 dB, Root Mean Squared Error (RMSE) of 0.0157, Percentage Root Mean Squared Difference (PRD) of 14.75%, and a Correlation Coefficient (CC) of 0.9983. By bridging the gap between consumer-grade and clinical-grade ECG monitoring, the proposed method sets a benchmark for real-time telemedicine and remote healthcare applications.