A self-attention 1D U-Net for radar-based contactless monitoring reduced log-spectral distance by 13.49 dB and improved Micro-F1 segmentation from 74.41% to 84.17%.
A self-attention 1D U-Net applied to radar-derived signals improves the fidelity of contactless phonocardiogram reconstruction and cardiac state segmentation.
Effect estimate: 13.4885 dB reduction
Absolute Event Rate: 84.17% vs 74.41%
Contactless vital signs monitoring is becoming increasingly relevant in scenarios where conventional sensors are impractical or not recommended. In this manuscript, a radar-based contactless system for the joint reconstruction of phonocardiogram (PCG) waveforms and cardiac state segmentation is illustrated. The proposed method exploits a self-attention one-dimensional (1D) U-Net fed by a pre-processed radar-derived input to estimate a PCG-like waveform, its envelope, and the four main cardiac phases: S1, systole, S2, and diastole. The accuracy of our method has been assessed on a public synchronized radar–PCG dataset acquired by means of a 24 GHz Doppler radar and a digital stethoscope. On the test subset, the proposed model achieved a 13.4885 dB reduction in log-spectral distance relative to the radar input signal, indicating a marked improvement in waveform fidelity. Segmentation performance also improved, with Micro-F1 increasing from 74.41% to 84.17% and Macro-F1 from 68.40% to 80.43% on average. Experimental results demonstrated the viability of real-time low-power embedded hardware deployment for contactless auscultation and continuous cardiac monitoring applications. The findings confirm that respiratory interference and low-amplitude signals complicate S2 detection, especially when exacerbated by subject motion.
Montanari et al. (Fri,) conducted a other in Cardiac monitoring. Self-attention 1D U-Net for radar-based PCG reconstruction vs. Baseline radar input signal was evaluated on Log-spectral distance reduction and segmentation performance (Micro-F1 and Macro-F1) (13.4885 dB reduction). A self-attention 1D U-Net for radar-based contactless monitoring reduced log-spectral distance by 13.49 dB and improved Micro-F1 segmentation from 74.41% to 84.17%.