Abstract
PURPOSE: Pathological myopia causes irreversible damage, yet early changes in high myopia patients with tessellated fundus are often overlooked. The identification of such patients relies on OCT and fundus imaging, but early detection is difficult. Therefore, we developed and validated machine learning models using lifestyle and clinical parameters. METHODS: Cross-sectional data were collected from high myopia patients with tessellated fundus at the First Affiliated Hospital of Guangxi Medical University (May 2023-July 2024). Data were split (7:3) into training and test sets. Five-fold cross-validation within the training set was used during LASSO regression to select key variables. Machine learning models, including Random Forest, Support Vector Machine, Linear Support Vector Machine, and XGBoost, were constructed. Model performance was evaluated by AUC, sensitivity, specificity, calibration curves, and decision curve analysis. RESULTS: 529 eyes were analyzed. Key modeling variables included education level, daily iPad usage time, daily reading time, daily reading time while lying down, lamp light source, daily outdoor activity time, spherical equivalent, axial length, corneal thickness, lens thickness, extent of PPA involvement, maximum radius of PPA, horizontal disc diameter, and disc tilt ratio. In the test set, the AUCs of the Random Forest, Support Vector Machine, Linear Support Vector Machine, and XGBoost models were 0.834, 0.815, 0.765, and 0.839, respectively. Both the training and test sets demonstrated good calibration and high clinical applicability. CONCLUSIONS: Machine learning models using lifestyle and clinical parameters can effectively screen for early pathological changes in high myopia with tessellated fundus, offering an auxiliary tool for clinicians.