Predicting diabetic retinopathy based on routine laboratory tests by machine learning algorithms.

Area of Science:

• Ophthalmology
• Endocrinology
• Data Science

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness in patients with type 2 diabetes mellitus (T2DM).
• Predictive models for DR risk stratification are crucial for timely intervention.
• Routine laboratory data offers a valuable, accessible resource for developing such models.

Purpose of the Study:

• To identify risk factors for DR in T2DM patients.
• To develop and validate machine learning (ML)-based predictive models for DR using routine laboratory data.
• To compare the performance of different ML algorithms for DR risk prediction.

Main Methods:

• Analysis of clinical data from 4259 T2DM inpatients.
• Development of a prediction model using the eXtreme Gradient Boosting (XGBoost) algorithm with 39 optimal variables.
• Comparison of XGBoost with Support Vector Machine (SVM), Gradient Boosting Decision Tree (GBDT), Neural Network (NN), and Logistic Regression (LR) models.
• Interpretation of the XGBoost model using the Shapley Additive exPlanation (SHAP) method.
• External validation of the best-performing model.

Main Results:

• Diabetic retinopathy (DR) was present in 47.69% of the T2DM patient cohort.
• The XGBoost model demonstrated superior performance with an AUC of 0.831, accuracy of 0.757, sensitivity of 0.754, specificity of 0.759, and F1-score of 0.752.
• SHAP analysis identified key risk factors contributing to DR development.
• External validation confirmed the model's predictive capability, achieving an accuracy of 0.650.

Conclusions:

• The XGBoost-based predictive model effectively assesses DR risk in T2DM patients using readily available laboratory data.
• This ML approach can assist clinicians in identifying high-risk individuals for DR.
• The model supports the implementation of personalized management strategies, particularly beneficial in resource-limited settings.