Heart rate estimation for U-Net and LSTM models combining multiple attention mechanisms

Abstract

Accurate heart rate (HR) monitoring is pivotal in modern medical technology, particularly for disease prevention and health management. This study proposes a novel HR estimation framework that leverages advanced deep learning techniques to accurately extract HR information from noisy photoplethysmography (PPG) signals. The proposed model, termed DRL-Unet, integrates a Denoising Autoencoder (DAE), U-Net architecture, and Long Short-Term Memory (LSTM) networks. To further enhance robustness and precision, the model incorporates a Multi-Head Attention mechanism and Residual Network (ResNet) modules. Comparative experiments demonstrate that DRL-Unet outperforms conventional deep learning models, achieving a Mean Absolute Error (MAE) of 1.69 bpm, Mean Squared Error (MSE) of 3.05 bpm², Root Mean Squared Error (RMSE) of 1.71 bpm, Mean Absolute Percentage Error (MAPE) of 1.15%, and Bias of 0.05 bpm. The model's effectiveness is validated on a public dataset from the IEEE Signal Processing Cup, confirming its superior performance under complex noise conditions. These findings highlight the potential of DRL-Unet to significantly improve the accuracy and reliability of HR estimation, thereby offering valuable advancements for early cardiovascular disease diagnosis and continuous health monitoring.