Decoding the night: a machine learning approach to diagnose and predict the severity of obstructive sleep apnea through clinical parameters

Obstructive sleep apnea (OSA) is a common sleep-related disorder associated with cardiovascular and metabolic comorbidities. Early diagnosis is essential; however, access to polysomnography is limited. Machine learning approaches may support clinical decision-making by predicting OSA using easily obtainable variables. This study aimed to develop an artificial neural network (ANN) model to predict the presence and severity of OSA. This multicenter study included adult patients who underwent diagnostic polysomnography at ten different sleep centers. Eighteen demographic and clinical parameters were used as input features, and four OSA severity categories served as output classes. The ANN architecture consisted of one hidden layer with 50 neurons (ReLU activation) and a Softmax output layer. The dataset was randomly divided into training (80%) and testing (20%) sets. A total of 1,954 participants (70.2% male) were included. Polysomnography classified 39.2% of participants as severe OSA, 29.4% as mild OSA, 15.0% as moderate OSA, and 16.4% as non-OSA. The ANN model demonstrated good discriminative performance across OSA severity categories, with class-specific AUC values ranging from 0.89 to 0.96. These findings support the potential utility of ML-based tools for screening and early risk stratification, particularly in resource-limited settings.