Non-Intrusive Electrocardiogram Reconstruction from Radar Signals via Deep Learning

Electrocardiogram (ECG) monitoring is crucial for cardiovascular health assessment. Traditional ECG systems require direct skin contact, which can be uncomfortable and impractical for continuous monitoring. This study proposes a radar-based ECG reconstruction approach using a Convolutional Recurrent Autoencoder (CRAE) network, using deep learning techniques to extract meaningful cardiac information from radar signals. Our method processes raw radar signals and reconstructs ECG waveforms with high fidelity by leveraging spatial and temporal dependencies through convolutional and bidirectional LSTM (Bi-LSTM)layers. The experimental results demonstrate strong agreement between the predicted and actual ECG signals, achieving a Pearson's correlation coefficient (PCC) of 0.889. The proposed deep learning-based method offers a promising noncontact alternative for cardiac monitoring, enhancing patient comfort and accessibility. Future work will explore improvements in feature extraction and clinical validation to further enhance reliability and applicability in real-world scenarios.