Comparative Evaluation of CNN Architectures for Wearable Heartbeat Detection Using Multimodal ECG and PPG Signals

Abstract

Wearable heartbeat monitoring has become increasingly important for continuous cardiovascular assessment in daily life. However, reliable heartbeat detection remains challenging under ambulatory conditions due to motion artefacts and signal degradation. Electrocardiography (ECG) provides reliable measurements of cardiac electrical activity, whereas photoplethysmography (PPG) offers a non-invasive and wearable-friendly alternative for continuous physiological monitoring. Although both signals have been widely used for heartbeat detection, most existing studies employ ECG or PPG independently, while limited attention has been given to multimodal physiological verification for recovering missed heartbeat detections. This paper presents a comparative evaluation of three one-dimensional convolutional neural network (1D-CNN) architectures, namely Baseline 1D-CNN, Multi-Scale CNN, and Residual 1D-CNN, for wearable heartbeat detection using synchronized ECG and PPG signals. ECG was employed as the primary heartbeat detection modality, while synchronized PPG signals were used for temporal verification to recover false negative heartbeat detections based on the physiological delay between ECG R-peaks and PPG pulse peaks. Experiments conducted on the PPG-DaLiA dataset using a subject-independent evaluation protocol showed that the Residual 1D-CNN achieved the best ECG-only performance, with sensitivity of 80.40% and an accuracy of 72.20%. After PPG-assisted temporal verification, the sensitivity and accuracy increased to 88.67% and 80.17%, respectively. These results show that PPG-assisted temporal verification consistently improves heartbeat detection performance and enhances the robustness of CNN-based wearable heartbeat detection under ambulatory monitoring conditions.