Kidney stones composition prediction using artificial intelligence ( KiSCAI) study

Abstract

Objectives To develop and validate explainable machine learning (ML) models predicting urinary stone composition from routinely available clinical variables and standardised morphological features, and to quantify the incremental value of morphology. Patients and Methods This retrospective cohort study included consecutive patients undergoing endourological treatment or spontaneous stone expulsion, with laboratory analysis showing a major component exceeding 50% of stone composition, between 2019 and 2024. The outcome was dominant stone composition, classified into five categories: calcium oxalate monohydrate (COM), calcium oxalate dihydrate (COD), calcium phosphate, uric acid, and cystine. Predictors included demographics, comorbidities, stone metrics, procedural details, and Daudon‐based morphological descriptors. Data were split into stratified training and validation cohorts (80% and 20%, respectively). Predictor stability was assessed using repeated‐resampling multiclass Least Absolute Shrinkage and Selection Operator (LASSO). Multiple supervised classifiers (logistic regression, support vector machine, random forest, ExtraTrees, Adaptive Boosting [AdaBoost] variants, Extreme Gradient Boosting [XGBoost], Categorical Boosting [CatBoost]) were fine‐tuned using GridSearch. Discrimination was assessed using macro‐averaged one‐vs‐rest area under the curve (AUC) and accuracy. Explainability relied on SHapley Additive exPlanations (SHAP). Results Among 442 patients (median age 51 years; 67% male), stone composition was COM in 41.0% ( n = 181), COD in 24.9% ( n = 110), calcium phosphate in 19.7% ( n = 87), uric acid in 10.9% ( n = 49), and cystine in 3.6% ( n = 16). LASSO stability highlighted reproducible predictors, including age, hereditary disease type, stone density, maximal diameter, and location. Clinical‐only models achieved good discrimination (macro‐AUC up to 0.809). Adding morphological features markedly improved performance, with ensemble models achieving excellent discrimination in validation (macro‐AUC up to 0.983 with CatBoost). Morphological variables ranked among the strongest contributors on SHAP. Conclusions A clinical‐only model can predict the major stone component and may support preoperative decision‐making. Performance further improved when morpho‐constitutional features were added, confirming Daudon's classification as the ground truth; however, this combined model relies on intra‐ and postoperative descriptors and is best regarded as an intra‐ or postoperative decision‐support tool rather than a preoperative one. External prospective validation is needed before clinical implementation.