Artificial intelligence-based prediction of metabolic dysfunction-associated steatotic liver disease (MASLD) and liver fibrosis in patients with obstructive sleep apnea: An innovative approach.

BACKGROUND: Obstructive sleep apnea (OSA) is strongly associated with metabolic dysfunction and may contribute to the development of hepatic disorders such as metabolic dysfunction-associated steatotic liver disease (MASLD) and liver fibrosis. Early identification of hepatic involvement in OSA is important for preventing long-term complications. This study aimed to develop and compare artificial intelligence (AI)-based models using non-invasive clinical parameters to predict steatosis, MASLD, and liver fibrosis in patients with OSA. METHODS: This prospective study included patients diagnosed with OSA based on polysomnography (PSG). Hepatic steatosis and fibrosis were assessed non-invasively using the controlled attenuation parameter (CAP) and liver stiffness measurement (LSM) obtained by FibroScan. A total of 700 patients were screened, and 517 patients who met the inclusion criteria were analyzed, including 42 demographic, anthropometric, laboratory, and sleep-related variables. A total of 42 variables were recorded, comprising 37 candidate predictors and 4 target outcomes. Following feature selection, 11 of the 37 candidates were retained for model training. Four supervised machine learning algorithms-logistic regression, support vector machine, random forest, and artificial neural network were trained to predict steatosis, MASLD, and liver fibrosis. Model performance was evaluated by accuracy, F1-score, and area under the ROC curve. Feature importance analysis identified the most influential predictors. RESULTS: Among the four models, the Random Forest algorithm achieved the best performance, with an accuracy of 96% for binary classification; in an exploratory five-class analysis, overall accuracy reached 92%, although predominantly driven by the majority F0 class. The most influential predictors were apnea-hypopnea index (AHI), minimum oxygen saturation (minSpO2), and oxygen desaturation index. A sensitivity analysis excluding elastography-derived features (CAP and LSM) was performed using 9 predictor variables (Model 2). Model 2 achieved best AUC values of 0.700, 0.706, and 0.634 for steatosis, MASLD, and fibrosis, respectively, compared with 0.967, 0.969, and 0.985 in the original 11-feature model (Model 1). CONCLUSION: AI-based models, particularly Random Forest, demonstrated high predictive performance in predicting MASLD and liver fibrosis in OSA patients. These models may support detection and risk stratification of hepatic involvement in sleep-disordered breathing.