An interpretable machine learning model for detecting vision-threatening diabetic retinopathy among patients with diabetic retinopathy: a web-based cross-sectional study.

The prevalence of vision-threatening diabetic retinopathy (VTDR) among enrolled diabetic retinopathy patients was 36.9%.

• Total enrolled patients: 1,124 T2DM patients with DR.
• Patients were categorized into VTDR and non-VTDR groups, with non-VTDR defined as mild-to-moderate non-proliferative diabetic retinopathy.
• Data were retrospectively extracted from electronic medical records at one hospital.
• The dataset was partitioned into training and testing sets at a 7:3 ratio.

Key factors associated with VTDR included diabetic treatment, T2DM duration, glycated hemoglobin levels, albuminuria, and anemia.

• These factors were identified as key associated features in the study population.
• SHAP (Shapley Additive Explanations) were used to interpret feature importance in the best-performing model.
• A simplified calculator was derived from the SHAP feature importance rankings and maintained strong diagnostic capacity.

The Support Vector Machine (SVM) model demonstrated superior performance compared to seven other machine learning models.

• SVM achieved an AUC of 0.879, accuracy of 0.837, precision of 0.833, Brier score of 0.129, and an F1 score of 0.756 in the testing cohort.
• The SVM model achieved the highest total score of 57 out of 64 using a comprehensive scoring system.
• Eight ML models in total were trained and evaluated using metrics including AUC, accuracy, and recall.
• Decision curve analysis and calibration curves confirmed the robustness and reliability of the models.

A web-based application was developed to demonstrate the potential clinical utility of the interpretable SVM model.

• The web-based application was designed to assist clinicians in prioritizing high-risk patients.
• The model uses routine clinical data as inputs, supporting cost-effective screening.
• The study authors describe the work as a 'proof-of-concept for a cost-effective screening tool'.
• External validation was noted as a requirement before broader clinical implementation.