Explainable Machine Learning-Based Prediction Model for Diabetic Nephropathy.

Area of Science:

• Metabolomics
• Machine Learning in Healthcare
• Nephrology Research

Background:

• Diabetic nephropathy (DN) is a major complication of diabetes, leading to kidney damage.
• Early detection and prediction of DN are crucial for effective management.
• Serum metabolites offer potential biomarkers for disease prediction.

Purpose of the Study:

• To investigate the association between serum metabolites and diabetic nephropathy.
• To develop and evaluate a machine learning model for predicting DN prevalence.
• To identify potential serum metabolite biomarkers for DN.

Main Methods:

• Utilized a dataset of 548 patients from the Second Affiliated Hospital of Dalian Medical University.
• Employed least absolute shrinkage and selection operator (LASSO) regression for feature selection, identifying 38 optimal features.
• Compared four machine learning algorithms (XGBoost, Random Forest, Decision Tree, Logistic Regression) using AUC-ROC, decision, and calibration curves.
• Applied Shapley Additive Explanation (SHAP) for quantifying feature importance and interactions.

Main Results:

• The extreme gradient boosting (XGB) model achieved the highest performance with an AUC value of 0.966 for DN screening.
• The XGB model demonstrated superior clinical net benefits and better fitting compared to other algorithms.
• Significant interactions were observed between serum metabolites and diabetes duration.
• Specific metabolites (C2, C5DC, Tyr, Ser, Met, C24, C4DC, Cys) were identified as important contributors to the predictive model.

Conclusions:

• An XGB-based machine learning model effectively screens for diabetic nephropathy.
• Identified serum metabolites, including C2, C5DC, Tyr, Ser, Met, C24, C4DC, and Cys, show promise as potential biomarkers for DN.
• The findings highlight the utility of metabolomics and machine learning in advancing DN prediction and management.