Cardiac Motion Modeling Based on Few-shot Learning and Radar

• 1、School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：With the advancement of technology and the popularization of medical knowledge among the public, people are becoming increasingly interested in their own heart movement information. Key information about heart movement is also of great significance in health care and disease warning. Based on the modeling of heart movement information from radar sensors, non-contact information related to the mechanical motion level of human heart during general physiological activities is obtained. With the help of deep neural networks’ strong data hiding feature representation and mining ability, the mechanical motion level and electrical signal conduction level of heart movement are fitted. At the same time, taking into account the difficulties and challenges in obtaining a large amount of high-quality radar data, a radar heart movement information modeling method based on small sample learning is proposed. This paper first uses 3D micro-focusing algorithm to process the human echo signal collected by radar, then adopts a small sample learning solution based on data enhancement to realize the nonlinear mapping from radar echo signal to electrocardiogram signal by proposing a method combining convolutional neural network and bidirectional long short-term memory. Establish a unified connection between the mechanical motion level of heart movement and the electrical signal conduction level of heart movement. A series of performance parameters were evaluated for the fitted electrocardiogram signal and the measured electrocardiogram signal, with an average value of 0.853 on the correlation coefficient index and an average absolute error of 0.161 millivolts. Finally, the effectiveness of the proposed algorithm and designed neural network was verified.

Keywords： biomedical engineering radar signal processing cardiac motion modeling few-shot learning