An enhanced machine learning algorithm for type 2 diabetes prognosis with a detailed examination of Key correlates

Area of Science:

• Medical Informatics
• Ophthalmology
• Machine Learning

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Accurate prediction and diagnosis models are crucial for timely intervention.
• Identifying key correlates can improve understanding of DR pathogenesis.

Purpose of the Study:

• To develop a high-performance prediction and diagnosis model for diabetic retinopathy (DR).
• To identify key correlates associated with DR using machine learning.
• To compare the performance of machine learning models against logistic regression.

Main Methods:

• Utilized a cross-sectional dataset of 3,000 patients.
• Employed recursive feature elimination cross-validation (RFECV) for feature selection.
• Developed and optimized four machine learning models: SVM, DT, RF, and GBDT.
• Used Shapley-additive explanations (SHAP) for factor analysis.

Main Results:

• The Gradient Boost Decision Tree (GBDT) model demonstrated superior performance (AUC: 0.8672).
• Identified six key correlates of DR: micronutrient protein/creatinine, 24-h micronutrient protein, fasting C-peptide, glycosylated hemoglobin, blood urea, and creatinine.
• The GBDT model achieved higher accuracy and identified fewer, more significant factors than logistic regression.

Conclusions:

• A superior, high-performance prediction model for DR was successfully constructed.
• The model effectively identified easily explainable key correlates of DR.
• Machine learning models offer improved prediction accuracy and efficiency compared to traditional statistical methods for DR.