Routine Urinary Biochemistry Does Not Accurately Predict Stone Type Nor Recurrence in Kidney Stone Formers: A...

Area of Science:

• Nephrology
• Biochemistry
• Data Science

Background:

• Urinary biochemistry aids in detecting and monitoring recurrent kidney stones.
• Currently, no predictive machine learning (ML) tools exist for kidney stone type or recurrence.
• Predictive modeling for kidney stones is crucial for personalized patient management.

Purpose of the Study:

• To build and validate ML models for predicting kidney stone type and recurrence.
• To utilize 24-hour urine biochemistry, age, gender, and stone composition as predictive features.
• To assess the efficacy of ML models in distinguishing stone types and predicting recurrence.

Main Methods:

• Data from three international cohorts (Southampton, Newcastle, Bern) were analyzed.
• ML models were developed for stone type (5 models) and recurrence (7 models) using UK data.
• Models were externally validated using Swiss data, employing complete case, multiple imputation, and oversampling techniques.

Main Results:

• An extreme gradient boosting (XGBoost) model effectively discriminated between calcium oxalate, calcium phosphate, and urate stones.
• This XGBoost model demonstrated strong performance on both internal and external validation.
• No developed ML models could accurately predict kidney stone recurrence.

Conclusions:

• 24-hour urinary biochemistry alone is insufficient for accurately predicting kidney stone recurrence.
• A single ML model successfully differentiated between various kidney stone types.
• Future research should incorporate radiomics and genomics for improved predictive tools for kidney stone recurrence.