A lightweight multi-feature fusion deep learning architecture for human ECG reconstruction from chest-worn accelerometer

Human electrocardiography (ECG) reconstruction from alternative human physiological signals provides a promising solution to overcome the inconvenience of direct electrode attachment. While previous approaches used generative models with photoplethysmography (PPG) and ballistocardiography (BCG), they suffered from high computational complexity and training instability. This article proposes a novel and lightweight deep learning architecture that reconstructs ECG from chest-worn accelerometer (ACC), which are more accessible and cost-effective. The proposed model integrates convolution-based (CNN) and graph-based (GNN) feature extraction modules to capture temporal patterns and inter-axis relationships. The best reconstruction result on the human physiological signal dataset achieved correlation coefficients of 0.9403 ± 0.0145 between reconstructed and ground truth ECG, outperforming CNN and GNN modules by 6.61% and 8.58%, respectively. The proposed architecture can potentially be used for a human ACC-to-ECG reconstruction approach.