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

Abstract

Objectives:

To assess the accuracy of machine learning models in predicting kidney stone composition using variables extracted from the electronic health record (EHR).

Materials and Methods:

We identified kidney stone patients (n = 1296) with both stone composition and 24-hour (24H) urine testing. We trained machine learning models (XGBoost [XG] and logistic regression [LR]) to predict stone composition using 24H urine data and EHR-derived demographic and comorbidity data. Models predicted either binary (calcium vs noncalcium stone) or multiclass (calcium oxalate, uric acid, hydroxyapatite, or other) stone types. We evaluated performance using area under the receiver operating curve (ROC-AUC) and accuracy and identified predictors for each task.

Results:

For discriminating binary stone composition, XG outperformed LR with higher accuracy (91% vs 71%) with ROC-AUC of 0.80 for both models. Top predictors used by these models were supersaturations of uric acid and calcium phosphate, and urinary ammonium. For multiclass classification, LR outperformed XG with higher accuracy (0.64 vs 0.56) and ROC-AUC (0.79 vs 0.59), and urine pH had the highest predictive utility. Overall, 24H urine analyte data contributed more to the models' predictions of stone composition than EHR-derived variables.

Conclusion:

Machine learning models can predict calcium stone composition. LR outperforms XG in multiclass stone classification. Demographic and comorbidity data are predictive of stone composition; however, including 24H urine data improves performance. Further optimization of performance could lead to earlier directed medical therapy for kidney stone patients.