Radar-based Non-contact Obstructive Apnea and Hypopnea Event Detection Using Quadrature-Temporal Point Clouds with Modified PointNet

Sleep apnea event detection using the radar sensor is attractive due to its non-contact fashion. In the radar baseband signal processing chain, the demodulation is a critical step affecting the quality of phase information involving the breathing signal, ultimately impacting the accuracy of detecting apnea-hypopnea events. However, some factors, such as phase wrapping and direct current (DC), lead to the distortion of the phase information and can hardly be obliterated. In this paper, by directly combining the quadrature baseband signals, a quadrature-temporal point cloud is proposed to represent the quadrature baseband signals involving breathing activity while avoiding the influences in the demodulation. In order to process the quadrature-temporal point cloud, a modified PointNet model is applied to classify the apnea-hypopnea event from the actual patients' nocturnal signals. A dataset from 31 obstructive sleep apnea-hypopnea syndrome (OSAHS) patients with simultaneously measured polysomnography (PSG) has been collected in a sleep center. Various experiments have been carried out to evaluate the performance. The overall accuracy and ablation experiments both show the superiority of the proposed method. Moreover, the proposed method achieves an average accuracy of 84% and 78% for subject-wise cross-validation and multi-class (apnea, hypopnea, and normal breathing) classification, respectively. All the experimental results demonstrate the effectiveness of the proposed method, showing potential applications for home sleep apnea testing (HSAT).