Machine Learning Models to Predict 24 Hour Urinary Abnormalities for Kidney Stone Disease.

Area of Science:

• Nephrology
• Medical Informatics
• Computational Biology

Background:

• Kidney stone disease management relies on accurate urine chemistry analysis.
• Predicting 24-hour urine abnormalities can guide empiric therapy and prevention strategies.
• Current methods for predicting urine abnormalities are limited.

Purpose of the Study:

• To assess the feasibility of using machine learning (ML) to predict 24-hour urine abnormalities.
• To compare the performance of ML models against traditional logistic regression.
• To identify key predictors for urine abnormalities in kidney stone disease.

Main Methods:

• Trained an XGBoost (XG) model and a logistic regression (LR) model using electronic health record data (n=1314).
• Developed an ensemble (EN) model combining XG and LR.
• Evaluated model performance using area under the receiver operating curve (AUC-ROC) for predicting urine volume, sodium, oxalate, calcium, uric acid, citrate, and pH.

Main Results:

• XGBoost demonstrated fair performance, comparable to logistic regression.
• The XG model excelled in predicting urine volume (AUC-ROC=0.59), uric acid (0.73), and elevated urine sodium (0.79).
• The ensemble model showed the best prediction for oxalate (0.70) and citrate (0.69). Key predictors included BMI, age, and gender.

Conclusions:

• Machine learning prediction of urine chemistry for kidney stone disease is feasible.
• ML models offer a potential tool to aid in guiding empiric therapy.
• Further model optimization could enhance dietary and pharmacologic prevention strategies.