A Metabolism-Based Interpretable Machine Learning Prediction Model for Diabetic Retinopathy Risk: A Cross-Sectional...

Area of Science:

• Metabolomics and Computational Biology
• Ophthalmology and Endocrinology

Background:

• Increasing global burden of diabetic retinopathy (DR), a leading cause of blindness.
• Limitations of current biomarkers for early detection and prediction of DR.
• Emerging evidence suggests metabolic alterations, including amino acids (AAs) and acylcarnitines (AcylCNs), in early-stage DR.

Purpose of the Study:

• To construct and validate a metabolite-based prediction model for DR risk in type 2 diabetes (T2D) patients.
• To identify key metabolic biomarkers for predicting DR development.
• To develop an interpretable model for clinical application in DR risk stratification.

Main Methods:

• Recruitment of T2D patients with and without DR.
• Construction of logistic regression and extreme gradient boosting (XGBoost) models using clinical features and metabolic profiles (AAs and AcylCNs).
• Model performance evaluation using discrimination (ROC AUC, PR AUC) and calibration (Brier score).
• Interpretation of the optimal model using Shapley Additive exPlanations (SHAP).

Main Results:

• The XGBoost model incorporating AAs and AcylCNs demonstrated superior predictive performance (ROC AUC = 0.82, PR AUC = 0.44, Brier score = 0.09).
• Specific metabolite thresholds associated with increased DR risk were identified: C18:1OH < 0.04 µmol/L, C18:1 < 0.70 µmol/L, threonine > 27.0 µmol/L, tyrosine < 36.0 µmol/L.
• Phenylalanine > 52.0 µmol/L was associated with a decreased risk of DR.

Conclusions:

• An interpretable XGBoost model utilizing AAs and AcylCNs effectively predicts DR risk in T2D patients.
• The identified metabolic biomarkers and their thresholds offer potential for early DR risk assessment.
• This approach can aid in identifying high-risk individuals for targeted interventions to prevent or delay DR onset.