Beyond Recurrent Models: Temporal Convolution for Robust Electrocardiogram Reconstruction from Cardio mechanical Signals

Unobtrusive cardiomechanical sensing modalities such as ballistocardiography (BCG) and seismocardiography (SCG) provide a promising pathway for continuous, electrode-free cardiac monitoring. However, reconstructing clinically interpretable electrocardiogram (ECG) signals from these mechanical measurements remains challenging due to noise, subject variability, and complex mechanical–electrical coupling. In this study, we propose a Bi-Temporal Convolutional Network (BiTCN) for robust cardiomechanical-to-ECG reconstruction. Unlike recurrent architectures that rely on hidden-state propagation, the proposed model uses multi-scale dilated temporal convolutions with gated residual updates, enabling explicit modeling of both local waveform morphology and long-range temporal context. The method is evaluated on three public datasets spanning distinct sensing conditions, including force-plate BCG, mattress-based optical-fiber BCG, and wearable SCG recordings. Across datasets, BiTCN improves waveform reconstruction fidelity compared to a CNN–BiLSTM baseline, increasing correlation from 0.577 to 0.603 and 0.778 to 0.807, while reducing RMSE from 0.748 to 0.722 and 0.453 to 0.426. Peak-level evaluation demonstrates improved beat recovery, with lag-tolerant R-peak detection increasing from 0.9667 to 0.9830 macro-F1 due to higher recall. Reconstructed ECG signals also enable accurate physiological estimation, achieving RR interval errors of 24–31 ms and heart-rate errors of 3.5–4.3 bpm. These results indicate that explicit multi-scale temporal convolution provides a stable and generalizable foundation for electrodeless cardiac monitoring.