Developing an explainable machine learning model using body composition to predict cardiovascular mortality in...

Area of Science:

• Nephrology
• Cardiology
• Artificial Intelligence

Background:

• Cardiovascular disease (CVD) is the primary cause of mortality in patients undergoing dialysis.
• Accurate prediction of CVD risk at the initiation of dialysis is currently limited.

Purpose of the Study:

• To develop and validate a machine learning model for predicting CVD-related mortality in patients starting dialysis.
• To integrate computed tomography (CT)-derived body composition features into the predictive model.

Main Methods:

• Trained and validated eight machine learning algorithms using clinical, laboratory, and CT-derived body composition data from incident dialysis patients.
• Employed feature selection techniques (logistic regression, LASSO) and evaluated models using discrimination, calibration, and decision curve analysis.
• Utilized Shapley Additive Explanations (SHAP) for model interpretability and developed a web-based risk calculator.

Main Results:

• Identified eight key predictors: age, diabetes, CVD history, cardiac intervention history, dialysis modality, skeletal muscle density, hemoglobin, and serum creatinine.
• The CatBoost model achieved an area under the receiver operating characteristic curve of 0.843 (internal validation) and 0.799 (external validation).
• SHAP analysis highlighted CVD, skeletal muscle density, and hemoglobin as significant contributors to mortality prediction.

Conclusions:

• An explainable machine learning model integrating CT-derived body composition effectively predicts CVD-related mortality in incident dialysis patients.
• This model offers potential for early risk stratification and personalized preventive interventions upon dialysis initiation.