تشخيص آپنه خواب از روي صداي ناي مبتني بر يادگيري عميق

Sleep apnea is a common sleep-related breathing disorder that can lead to serious cardiovascular an‎d cognitive consequences an‎d reduced quality of life. Although polysomnography provides high diagnostic accuracy, its cost an‎d complexity have driven interest in low-cost methods for screening an‎d supporting the diagnostic process. This thesis addresses audio-based respiratory event detection, with the goal of identifying short audio segments an‎d assigning binary labels (apnea vs. non-apnea). In other words, the focus is on segment-level event detection rather than making a final decision about whether an individual has sleep apnea.To achieve this goal, two complementary approaches are proposed an‎d eva‎luated: (1) a primary representation-learning method based on the Siamese-M2D architecture with decision-making in an embedding space, an‎d (2) an alternative direct-classification method using Anisotropic-ResNet18 with multi-view time–frequency inputs. Experimental eva‎luation was conducted on two datasets: an in-house tracheal sound dataset recorded in a clinical setting an‎d the public PSG-Audio dataset. Three scenarios were designed, including within-dataset eva‎luation on each dataset an‎d a cross-dataset eva‎luation to examine generalizability. The results showed that on the in-house dataset, Siamese-M2D (with an appropriate decision configuration) achieved the best balance between precision an‎d recall (F1-score = 0.80), whereas on the public dataset, Anisotropic-ResNet18 achieved superior performance (F1-score = 0.86). In the cross-dataset scenario, the observed performance dro‎p indicated a distribution shift due to differences in recording conditions, environmental noise, an‎d event labeling protocols; nevertheless, certain model configurations were found to be suitable for screening purposes (high recall an‎d reduced false negatives). Overall, the findings suggest that audio-based approaches, despite cross-domain generalization challenges, have strong potential for developing low-cost, non-invasive tools for screening an‎d assisting in the detection of sleep-related respiratory events.

زينب زالي

فرزانه شايق بروجني

عليرضا بصيري , هدي صفائي پور