[Development of an OCTA-based risk prediction model for retinopathy in patients with obstructive sleep apnea and...

Objective:To identify risk factors for retinopathy in patients with obstructive sleep apnea（OSA） and hypertension, and develop a machine learning-based risk prediction model by integrating optical coherence tomography angiography（OCTA） -derived retinal microvascular parameters. Methods:This study enrolled 98 adult patients diagnosed with OSA and hypertension at China-Japan Friendship Hospital and Capital Medical University afflicted Beijing Tiantan Hospital, from June to December 2024. Clinical characteristics, polysomnography（PSG） indices, and OCTA parameters were collected as input features. Five machine learning models-logistic regression, decision tree（DT）, random forest（RF）, support vector machine（SVM）, and XGBoost-were developed and evaluated using accuracy, F1-score, and area under the ROC curve（AUC）. Results:Compared with the non-retinopathy group, the retinopathy group exhibited significantly elevated apnea-hypopnea index（AHI）, oxygen desaturation index（ODI）, and reduced minimum SpO2, alongside impaired OCTA-VLD（all P<0.05）. Among the models, XGBoost demonstrated superior predictive performance. The comparative analysis of machine learning models revealed that the XGBoost classifier achieved optimal diagnostic performance, with an accuracy of 0.774, F1-score of 0.588, and a superior AUC of 0.940, significantly outperforming logistic regression（AUC: 0.813）, decision tree（AUC: 0.797）, random forest（AUC: 0.893）, and support vector machine（AUC: 0.843）. SHAP interpretability analysis identified VLD, AHI, and minimum SpO2 as dominant predictors, collectively contributing 68.3% to the models decision-making. Conclusion:The XGBoost-based multimodal model integrating OCTA and PSG biomarkers enables precise risk stratification for retinopathy in OSA-hypertension comorbidity. This non-invasive framework elucidates the synergistic interplay between microvascular dysfunction and nocturnal hypoxemia, providing a quantifiable tool for early risk prediction of target organ damage. Its clinical translation potential lies in guiding personalized interventions to mitigate systemic complications in high-risk populations.