A Hybrid Deep Learning Model to Detect Obstructive Sleep Apnea for Adult Patients in Bangladesh

Obstructive Sleep Apnea (OSA) represents the most common form of sleep-disordered breathing and is associated with serious and long-term health consequences. Despite its prevalence, OSA remains substantially underdiagnosed in developing countries such as Bangladesh, primarily due to limited public awareness and the prohibitive cost of standard diagnostic procedures like polysomnography (PSG). To address this diagnostic gap, we propose a novel hybrid deep learning framework for the automated classification of OSA severity. The model leverages nasal airflow and blood oxygen saturation (SpO2) signals collected from Bangladeshi patients, with preprocessing techniques including filtering and temporal normalization to enhance signal quality. The processed data is segmented into 5 minute windows for feature extraction and classification into four severity categories-Normal, Mild, Moderate, and Severe-based on the Apnea-Hypopnea Index (AHI). The proposed model achieves a validation accuracy of 93% and a micro-averaged area under the ROC curve (AUC) of 0.99. Notably, the model exhibits strong performance in detecting both Normal and Severe cases, achieving a precision of 1.00 and recall of 0.97 for the Normal class, and a precision of 0.80 and recall of 0.76 for the Severe class. This study contributes the first application of a deep learningbased OSA detection system tailored to Bangladeshi patient data, offering a cost-effective and scalable alternative to expensive polysomnography.