Machine Learning Prediction of Kidney Stone Composition Using Electronic Health Record-Derived Features

Area of Science:

• Nephrology
• Biomedical Informatics
• Data Science

Background:

• Kidney stone composition is crucial for targeted treatment.
• Predicting stone type aids in developing personalized medical therapies.
• Electronic Health Records (EHR) offer rich data for predictive modeling.

Purpose of the Study:

• To evaluate machine learning models for predicting kidney stone composition.
• To compare model performance using EHR data versus 24-hour urine testing.
• To identify key predictors of kidney stone types.

Main Methods:

• Trained XGBoost and logistic regression models on 1296 kidney stone patients.
• Utilized 24-hour urine data, demographic, and comorbidity data from EHR.
• Evaluated models for binary (calcium vs. non-calcium) and multiclass stone prediction using ROC-AUC and accuracy.

Main Results:

• XGBoost achieved 91% accuracy in binary classification; logistic regression excelled in multiclass (64% accuracy).
• Uric acid and calcium phosphate supersaturations, and urinary ammonium were key predictors for binary classification.
• Urine pH was the primary predictor for multiclass stone classification.
• 24-hour urine analyte data significantly improved model performance over EHR data alone.

Conclusions:

• Machine learning models can effectively predict kidney stone composition, particularly calcium stones.
• Logistic regression demonstrates superior performance in multiclass stone classification.
• Integrating 24-hour urine data with EHR information enhances predictive accuracy, paving the way for earlier, targeted patient therapies.