Application of Machine Learning to Predict Symptomatic Recurrence Events for Kidney Stone Patients.

Introduction: Kidney stone recurrence can be reduced by implementing AUA medical management guidelines. We assessed if machine learning (ML) could identify patients at risk for symptomatic kidney stone recurrence events.

Methods: We retrospectively reviewed electronic health records (EHR) with kidney stone diagnosis over a 16-year period (January 2008 to December 2023). Using historical data from a large integrated health system, we applied supervised machine learning to build a model that identifies patients at risk for symptomatic recurrence events within 12 months following an initial stone encounter with a urologist. The model used 952 candidate features drawn from both a clinician-curated set of kidney stone specific factors and a general set of common diagnoses, laboratory results, medications, procedures, and utilization records were used as inputs to the model.

Results: Our model was tested and trained on data collected for 154,876 urinary stone patients over 16 years. 1,439,671 unique kidney stone encounters were attributable to this population. The algorithm was trained on 123,900 (80%) and tested on 30,976 (20%) patients. In the test set, the model predicted 1-year risk of symptomatic recurrence with an area under the receiver operating characteristic curve (AUROC) of 0.727.

Conclusion: Machine learning models can effectively discriminate between high and low risk of urinary stone recurrence events.