Abstract
BACKGROUND: We developed an artificial intelligence (AI) model to predict the severity of obstructive sleep apnea syndrome (OSAS). METHODS: We used data from 750 inpatients at a research hospital between 2021 and 2023. The dataset comprises 20 attributes, including demographic information, medical history, anthropometric measurements, and polysomnography (PSG) data. The target attribute was the apnea-hypopnea Index (AHI), from which OSAS severity was determined. Data preprocessing included min-max scaling for normalization and the Synthetic Minority Over-sampling Technique algorithm to address the class imbalance, increasing the dataset size to 1250. We invented and further developed a multilayer artificial neural network (ANN) model to predict OSAS severity and evaluated its performance using k-fold cross-validation. We also performed an information gain analysis to rank the features by importance. RESULTS: The ANN model accurately predicted OSAS severity (area under the receiver operating characteristic curve: 0.966, CA: 0.880). Information gain analysis revealed strong associations between OSAS severity and the Epworth Sleepiness Scale, lowest nighttime oxygen saturation, percentage of sleep time with oxygen saturation between 80% and 90% during the night, and neck thickness. These identified features represent important risk factors for early OSAS diagnosis and treatment. CONCLUSION: Our findings suggest that AI-based models can effectively predict OSAS severity. This research may contribute to the development of next-generation diagnostic tools for OSAS diagnosis and risk assessment. AI can readily determine OSAS severity from overnight pulse oximetry recordings, combined with other risk factors, in patients with suspected OSAS.