Abstract
OBJECTIVES: To investigate the role of retinal characteristics for high myopia (HM) prediction based on multi-machine learning (ML) and to provide an interpretable framework for the results. METHODS: A total of 2981 patients (2981 eyes) were included, comprising 1191 HM eyes and 1790 non-HM eyes. A deep semantic segmentation network was used to quantify retinal structural parameters. Five ML algorithms were evaluated to develop predictive models, and SHapley Additive exPlanations method was applied to analyze feature contribution to the outcomes. RESULTS: The eXtreme Gradient Boosting achieved an accuracy of 0.81 (95% confidence interval [CI] 0.78-0.85), and an area under the receiver operating characteristic curve of 0.87 (95% CI 0.84-0.89), outperforming other models. The 12 most important factors affecting prediction included tessellation density, seven vascular parameters, two parapapillary atrophy parameters, and two optic disc parameters. The tessellated density >0.025, width of parapapillary atrophy >400 um, parapapillary atrophy area >0.60 × 10(6) um(2) were associated with an increased risk of HM. The mean curvature of the arteries >0.00063, diameter of vessels >55.2 um, curvature of the veins >0.00128, vertical diameter of the optic disc >1320 um, diameter of veins >58.5 um, and diameter of artery >47.1 um was associated with a decreased risk of HM. CONCLUSION: The XGBoost model outperformed other algorithms, and SHAP-derived cut-off values for critical risk factors enhanced clinical interpretability.