Deep Learning Fused Wearable Pressure and PPG Data for Accurate Heart Rate Monitoring

Photoplethysmography (PPG) provides a non-invasive method to detect heart rate but, due to inherent noise, fails to accurately and reliably capture the true heartbeat waveform and phase needed by many cardiac activity monitoring applications such as measuring heart rate variability and blood pressure. In this work, we contribute 1) a multimodal heart rate and phase sensing device which is capable of capturing data from 17 channels including PPG, Accelerometer and Gyro at a high sampling rate of 500Hz, with a new pressure-sensing channel at 80Hz; and 2) a deep learning model to fuse multi-channel data to derive heart rate (HR) with high accuracy compared to ground truth from a reference electrocardiography (ECG) signal. We showed the sensors', channels' and locations' contribution separately toward heart rate measurement and identified a preferred site for HR detection. We demonstrated the improved HR measurement on 21 healthy participants under three different activities, namely, stationary, walking and running. Our system achieved a stationary average absolute error (AAE) of 0.47bpm, 0.79bpm (walking) and 0.89bpm (running). The respective single heartbeat standard deviations were 28.43ms, 40.3ms and 34.14ms relative to an electrocardiography (ECG) detected R-peak. To the best of our knowledge, we are the first to show that attachment pressure changes of a PPG sensor improved measurement accuracy by 8.0ms (25.6%).

Automated summary

This study introduces a multimodal heart rate and phase sensing device, along with a deep learning model that can accurately derive heart rate by fusing multi-channel data. Experimental results demonstrate improved accuracy in heart rate measurements under different activities like stationary, walking, and running.