A study on non-contact ECG signal reconstruction based on MST-ECGNet deep learning model
GUO Tianjiao
YUAN Nianzeng
AN Qiang
LYU Hao
WANG Jingzhe
ZHANG Zhiyuan
LONG Yunuo
LIU Zhenhua
XUE Huijun
Abstract：Objective To address the limitations of conventional electrocardiogram(ECG)monitoring technologies,including discomfort and inability for continuous monitoring induced by electrodes,this paper proposed an accurate non-contact ECG signal reconstruction method by combining bio-radar and deep learning technology.Methods First,a 94 GHz continuous-wave bio-radar was employed to capture thoracic micro-motion signals,from which cardiac mechanical motion signals were extracted using variational mode decomposition.Subsequently,a multi-scale transformer for ECG reconstruction network(MST-ECGNet)was constructed,employing a dual-path feature extraction architecture with the multi-scale cardiac network to extract local multi-scale features and Transformer-Encoder to extract the global temporal dependencies.After feature fusion,cardiac mechanical motion signals were translated into ECG signals by Transformer-Decoder.Results Experimental results demonstrated that the reconstructed ECG signals exhibited a Pearson correlation coefficient of 0.956 with reference signals,indicating high consistency.Compared to existing non-contact methods,the proposed approach showed superior performance in waveform reconstruction accuracy.Conclusion This paper proposes a non-contact ECG signal reconstruction method based on the MST-ECGNet model.This model takes into account both multi-scale local and global feature extraction,achieving precise ECG signal reconstruction and providing a non-contact,long-term,and dynamic solution for the diagnosis and monitoring of cardiovascular diseases.
Keywords：bio-radarnon-contactneural networksmulti-scale featureencoder-decoder networkelectrocardiogramcardiac mechanical motioncardiovascular diseases
Publication Date：2025-10-31
Online Publishing Date：2026-08-26(First online date of this platform, not the publication date of the document)
Pages：7( 1353-1358,1363 )