Predicting myopia risk using a machine learning model based on fundus imageomics.

Area of Science:

• Ophthalmology and Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Myopia is a growing global health concern, particularly in school-aged children.
• Early prediction of myopia risk is crucial for timely intervention and management.
• Quantitative analysis of retinal features from color fundus photography (CFP) offers potential for risk assessment.

Purpose of the Study:

• To develop a machine learning (ML) model using CFP data to predict myopia risk in children.
• To identify key retinal imageomics features associated with myopia progression.
• To evaluate the predictive performance of various ML algorithms for myopia risk stratification.

Main Methods:

• A cross-sectional study involving 2,184 children aged 6-10 years with CFP data.
• Extraction of 146 retinal imageomics features using the EVisionAI platform, alongside age and sex.
• Feature selection via LASSO regression and expert review, followed by model construction using RF, XGBoost, and LightGBM.

Main Results:

• The Random Forest (RF) model achieved the highest predictive performance (AUC = 0.798), outperforming LightGBM and XGBoost.
• Key predictors identified include age, nasal disc-foveal distance, atrophic areas, and vascular parameters.
• The RF model demonstrated high specificity (0.80) and moderate sensitivity (0.59), with strong calibration.

Conclusions:

• Quantitative CFP-derived imageomics combined with ML can effectively predict myopia risk in school-aged children.
• The developed RF model, utilizing age, retinal distances, and vascular features, shows significant clinical value.
• This approach offers a promising tool for early myopia risk stratification and personalized management strategies.