SHAP Analysis of Machine Learning-Based Carotid Atherosclerosis Prediction of Older Adults with OSA, a Multicenter Study.

Background: Carotid atherosclerosis (CAS) is a critical cardiovascular complication in elderly patients with obstructive sleep apnea (OSA). Current risk assessment tools inadequately capture OSA-specific pathophysiological mechanisms for CAS prediction in this high-risk population. Methods: This multicenter retrospective study included 1196 elderly patients (≥60 years) with polysomnography-confirmed OSA from six tertiary hospitals in China. CAS was diagnosed by carotid ultrasound. LASSO (Least Absolute Shrinkage and Selection Operator) regression identified optimal predictive features from 18 candidate variables. Four machine learning algorithms were developed and validated using 5-fold cross-validation. Model interpretability was achieved through SHapley Additive exPlanations (SHAP) analysis. Results: Among participants, 273 (22.8%) had CAS. LASSO regression selected eight optimal features. XGBoost achieved the best performance with test AUC of 0.854, accuracy of 79.8%, sensitivity of 81.2%, and specificity of 78.5%. SHAP analysis revealed systolic blood pressure (importance: 0.3148) and percentage of sleep time with oxygen saturation <90% (T90, importance: 0.2660) as the most influential predictors, surpassing traditional apnea-hypopnea index. Other significant predictors included alcohol consumption, mean oxygen saturation, body mass index, age, platelet count, and smoking status. Conclusion: This study developed the first machine learning-based CAS prediction model for elderly OSA patients, achieving clinically relevant performance (AUC=0.854). The prominence of T90 over conventional apnea-hypopnea index suggests nocturnal hypoxemic burden is more important than respiratory event frequency for cardiovascular risk stratification in this population.