Contactless Cardiac Detection of Arrhythmia Based on Distribution Encoding Network With Microwave Biomedical Radar

As microwave sensing for cardiac monitoring advances, characterizing the population-level distributions of microwave cardiogram (MCG) and achieving robust statistical consensus are increasingly important. This article proposes a distribution-encoding technique based on dimensionality reduction that enables heatmap visualization and provides intuitive insight into the distribution of abnormal cardiac patterns. Four typical classes of arrhythmias are presented, with a particular focus on two arrhythmias: atrial premature beat (APB) and atrial fibrillation (AF). To leverage prior knowledge in distribution coordinates while automatically learning discriminative features, a distribution encoding neural network has been proposed. Validation on data from 64 patients (6115 samples) collected at two hospitals shows that both 24GHz continuous-wave and 60GHz frequency-modulated continuous-wave radars can effectively discriminate arrhythmias with the proposed learning mechanism. In a three-class classification (AF, APB, normal), accuracy reaches 96.15% on a small dataset of 257 samples and 90.99% on a larger dataset of 2964 samples. For common three-, four-, and five-class classifications, overall accuracies are approximately 80%, 75%, and 70%, respectively. Ablation experiments are presented to analyze the effectiveness of the learning mechanism.