Video-Based Contactless Detection of Sleep Apnea With Deep-Learning Model

In recent years, sleep apnea (SA) detection has attracted increasing attention worldwide. Polysomnography (PSG) is currently considered the gold standard for sleep monitoring, but its limitations have hindered widespread adoption. Therefore, we propose a video-based contactless SA detection technique. Only a single camera is needed to capture sleep images, and images of the nose, throat, and face are selected to extract respiratory motion signals and physiological signals, including optical flow, remote photoplethysmography (rPPG), and heart rate variability (HRV). These signals are used for SA classification on per-minute segments and to assess the severity of the apnea-hypopnea index (AHI) throughout overnight sleep. Compared with existing methods, our proposed approach has the advantages of being noncontact, low cost, and easy to use. This allows users to maintain their usual sleep conditions, making monitoring more convenient. In evaluation experiments using a self-constructed rPPG-SA dataset involving 61 participants, the proposed method achieves results comparable to existing contact-based methods with ten subjects as training data and the other 51 subjects as testing data. The accuracy for binary classification of SA segments reaches 78.45% with a sensitivity of 79.03% and a specificity of 78.12%. In further experiments predicting the severity of AHI over the entire night, the accuracy for detecting moderate and severe levels (AHI
$\geq 15$
) reaches 76.47% with a sensitivity of 84.85% and a specificity of 61.11%. Therefore, the proposed method not only offers the advantages of convenience and comfort in terms of device usage but also achieves comparable performance to contact-based algorithms.