Interpretable machine learning algorithms for diagnostic prediction of diabetic retinopathy

Area of Science:

• Ophthalmology and Medical AI
• Machine Learning for Healthcare Diagnostics

Background:

• Diabetic Retinopathy (DR) is a leading cause of blindness, requiring early detection.
• Traditional DR screening relies on manual evaluation, while AI offers scalable solutions.
• Clinical trust in AI diagnostics is hindered by a lack of transparency.

Purpose of the Study:

• To develop a standardized, interpretable machine learning (ML) framework for DR risk prediction.
• To enhance diagnostic efficiency and accuracy by integrating model interpretability with performance.
• To bridge the gap between advanced AI and clinical applicability in DR screening.

Main Methods:

• Evaluated eleven ML algorithms with hyperparameter optimization and cross-validation.
• Developed a dynamic weighted voting ensemble (Voting_soft) integrating multiple classifiers.
• Utilized SHAP (Shapley Additive exPlanations) for feature interpretability analysis.

Main Results:

• Identified fourteen optimal clinical predictors for DR using LightGBM and AUC analysis.
• The dynamic weighted voting ensemble (Voting_soft) outperformed individual ML models in AUC.
• SHAP analysis highlighted age, triglycerides, sex, and HDL-C as key DR predictors.

Conclusions:

• Presents a novel ML-based DR risk prediction system with high accuracy and interpretability.
• SHAP analysis provides clinicians with actionable insights, enhancing diagnostic decision-making.
• The dynamic voting ensemble sets a new standard for interpretable multi-model integration in medical diagnostics.