A Non-Contact ECG Reconstruction Method Based on FMCW Radar and Adaptive VMD-Assisted Deep Learning

To address the limitations of conventional ECG monitoring, including the need for direct skin contact, discomfort during long-term use, and limited applicability in special monitoring scenarios, this paper proposes a non-contact ECG reconstruction method based on FMCW radar and deep learning. In the signal preprocessing stage, adaptive Variational Mode Decomposition (VMD) optimized by Ant Colony Optimization (ACO) is introduced to suppress respiratory harmonic interference and improve heartbeat signal separation. In the reconstruction stage, a deep learning framework integrating convolutional neural networks, multi-head attention, and LSTM is developed to capture local temporal features and global sequential dependencies, enabling the mapping from radar-derived heartbeat signals to ECG waveforms. Experiments on a public dataset and additional hospital-collected data demonstrate that the proposed method can reconstruct ECG signals with high waveform consistency relative to reference ECGs. Quantitative results show that the reconstructed signals achieve low RMSE, high PCC, and good RR-interval agreement with the ground-truth ECG. The results indicate the feasibility of the proposed method for non-contact ECG reconstruction and suggest its potential for continuous physiological monitoring.