Routine Urinary Biochemistry Does Not Accurately Predict Stone Type Nor Recurrence in Kidney Stone Formers: A Multicentre, Multimodel, Externally Validated Machine-Learning Study

Objectives: Urinary biochemistry is used to detect and monitor conditions associated with recurrent kidney stones. There are no predictive machine learning (ML) tools for kidney stone type or recurrence. We therefore aimed to build and validate ML models for these outcomes using age, gender, 24-hour urine biochemistry, and stone composition. Materials and Methods: Data from three cohorts were used, Southampton, United Kingdom (n=3013), Newcastle, United Kingdom (n=5984), and Bern, Switzerland (n=794). Of these 3130 had available 24-hour urine biochemistry measurements (calcium, oxalate, urate [Ur], pH, volume), and 1684 had clinical data on kidney stone recurrence. Predictive ML models were built for stone type (n=5 models) and recurrence (n=7 models) using the UK data, and externally validated with the Swiss data. Three sets of models were built using complete cases, multiple imputation, and oversampling techniques. Results: For kidney stone type one model (extreme gradient boosting [XGBoost] built using oversampled data) was able to effectively discriminate between calcium oxalate, calcium phosphate, and Ur on both internal and external validation. For stone recurrence, none of the models were able to discriminate between recurrent and nonrecurrent stone formers. Conclusions: Kidney stone recurrence cannot be accurately predicted using modeling tools built using specific 24-hour urinary biochemistry values alone. A single model was able to differentiate between stone types. Further studies to delineate accurate predictive tools should be undertaken using both known and novel risk factors, including radiomics and genomics.