Apnea/Hypopnea Detection Using Modified YOLOv8n Model on Non-EEG Signals

Sleep apnea is a sleep breathing disorder, causing ineffective sleep, that may lead to serious problems. Severity of sleep apnea is determined by an apnea-hypopnea index (AHI) that is usually obtained manually with both labor intensive and time consuming. Additionally, while the popularity of home sleep testing (HST) is increasing, many HST devices still lack EEG monitoring. Thus, our goal is to develop a machine learning model for detecting apnea and hypopnea events using 3 non-EEG signals to determine a severity level of sleep apnea for a full overnight record. We propose a modified YOLOv8n model for this task. The proposed model was trained, validated, and tested using 2 datasets including 2,056 and 23 full overnight records from the Multi-Ethnic Study of Atherosclerosis (MESA) and the St. Vincent's University Hospital/University College Dublin Sleep Apnea (UCD), respectively. Testing performances of the proposed model were evaluated using 15% record-based data splitting for MESA and 23-fold cross validation for UCD. By comparing with selected existing models, the proposed model achieves reasonably good sample-based testing performances. To assess more realistic testing performances, the severity levels of sleep apnea were predicted using the detected apnea/hypopnea events with and without groundtruth total sleep times (TSTs). The results show that the proposed model provides severity-level (event-based) performances substantially better than the existing model. The proposed model also achieves moderate to substantial agreement with the groundtruth severity levels where the Cohen's kappa coefficients are between 0.61 and 0.81.