Multiscale temporal deep learning for ECG waveform estimation from contactless radar signals

Electrocardiography (ECG) is widely used for monitoring cardiac activity and supporting the assessment of cardiovascular conditions. Conventional ECG systems require skin-contact electrodes, which may cause discomfort and limit their suitability for prolonged monitoring. Contactless radar sensing provides a non-contact approach for capturing chest-wall movements associated with cardiac activity. However, ECG waveform estimation from radar signals is challenging because radar measures mechanical chest-wall motion, whereas ECG represents electrical cardiac activity. In addition, respiratory interference, motion artifacts, phase distortion, and inter-subject variability can affect the relationship between radar and ECG signals. This study presents a multiscale temporal deep learning framework for ECG waveform estimation from contactless radar signals. The proposed framework combines radar signal preprocessing, feature extraction, multiscale residual convolutional learning, bidirectional long short-term memory (BiLSTM) networks, and multi-head attention to learn temporal and local signal representations from radar-derived features. Radar in-phase and quadrature (I/Q) signals are processed through phase extraction, denoising, and feature extraction to construct the input representation for ECG estimation. The framework was evaluated using a publicly available synchronized radar-ECG dataset containing 30 subjects under the resting scenario. A subject-independent evaluation protocol was used, with separate participants assigned to training, validation, and testing subsets. Across five repeated subject-independent randomized splits, the proposed framework achieved a mean absolute error (MAE) of 0.07584, a root mean square error (RMSE) of 0.14261 and Pearson correlation coefficient (Pearson r) of 0.64459. These results demonstrate the feasibility of estimating the overall temporal structure of ECG waveforms from contactless radar measurements under resting conditions.