Construction and validation of a urinary stone composition prediction model based on machine learning.

Area of Science:

• Urology
• Medical Informatics
• Machine Learning

Background:

• Accurate preoperative urinary stone composition is crucial for personalized surgical planning.
• Current methods often lack pre-operative compositional data, hindering optimal treatment selection.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting urinary calculi composition preoperatively.
• To assess the clinical utility of the ML model in guiding surgical strategies.

Main Methods:

• Retrospective cohort study of 708 patients with urinary calculi.
• Feature selection using LASSO regression and logistic regression.
• Model development and validation using AUC and SHAP values for interpretability.

Main Results:

• Developed distinct ML models for calcium oxalate, infection, uric acid, and calcium-containing stones.
• Achieved high AUC values (0.845-0.961) for predicting stone composition.
• Identified key predictors including CT value, urinary oxalate/calcium, stone size, and urinary pH.

Conclusions:

• An integrated ML model accurately predicts preoperative urinary stone composition.
• Combining imaging features with metabolic indicators improves predictive performance.
• The model supports personalized surgical strategies by providing pre-operative stone analysis.