Retinal features as predictive indicators for high myopia: insights from explainable multi-machine learning models.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• High myopia (HM) poses a significant risk for vision-threatening conditions.
• Accurate prediction of HM is crucial for early intervention and management.
• Retinal structural characteristics are increasingly recognized as potential biomarkers for HM.

Purpose of the Study:

• To evaluate the efficacy of multiple machine learning (ML) algorithms in predicting high myopia (HM) based on retinal features.
• To develop an interpretable framework for understanding the contribution of retinal parameters to HM prediction.

Main Methods:

• A deep semantic segmentation network was employed to extract quantitative retinal structural parameters from 2981 patient eyes (1191 HM, 1790 non-HM).
• Five distinct ML algorithms were trained and evaluated for their predictive performance.
• The SHapley Additive exPlanations (SHAP) method was utilized to analyze feature importance and model interpretability.

Main Results:

• The eXtreme Gradient Boosting (XGBoost) model demonstrated superior performance, achieving an accuracy of 0.81 and an AUC of 0.87.
• Twelve critical features were identified, including tessellation density, vascular parameters, parapapillary atrophy characteristics, and optic disc measurements.
• Specific thresholds for tessellated density, parapapillary atrophy width/area, and various vascular/optic disc parameters were associated with increased or decreased HM risk.

Conclusions:

• The XGBoost model, leveraging retinal characteristics, effectively predicts high myopia.
• SHAP analysis provided crucial insights into the predictive power of specific retinal features, enhancing clinical applicability.