Machine learning predicting myopic regression after corneal refractive surgery using preoperative data and fundus...

Area of Science:

• Ophthalmology
• Medical Artificial Intelligence
• Biomedical Data Science

Background:

• Myopic regression is a common complication after refractive surgery.
• Predicting myopic regression typically requires long-term patient observation.
• Identifying high-risk patients preoperatively is crucial for surgical planning.

Purpose of the Study:

• To develop and validate machine learning models for identifying patients at high risk of myopic regression.
• To utilize preoperative data and fundus photography for early prediction of refractive regression.
• To assess the predictive performance of integrated machine learning models.

Main Methods:

• Retrospective study with chronological data splitting into training (1606 eyes) and validation (403 eyes) sets.
• Development of ResNet50 (image analysis) and XGBoost (data integration) machine learning models.
• Evaluation of predictive performance for myopic regression (>0.5 D) at 4-year postoperative follow-up.

Main Results:

• The combined machine learning model integrating all factors and fundus photography achieved an ROC-AUC of 0.753.
• This integrated model outperformed a ResNet50-only model (ROC-AUC, 0.673).
• Key predictive features included fundus photography, anterior chamber depth, planned ablation thickness, age, and central corneal thickness.

Conclusions:

• Machine learning offers an efficient, labor-free strategy for preoperative risk assessment of myopic regression.
• This tool can assist surgeons in patient counseling and surgical option selection.
• Early identification of high-risk patients can improve refractive surgery outcomes.