Investigation of the multifactorial regulatory mechanisms underlying compositional changes in recurrent urinary stones and development of a machine learning-based personalized predictive model.

To characterize compositional changes in recurrent urinary calculi, identify associated factors, and develop/validate a prediction model to support risk assessment and individualized prevention. A retrospective cohort compared initial vs. recurrent stone composition and transformation patterns. Patients were classified as composition-unchanged (n = 113) or composition-changed (n = 89). Group differences were tested, and predictors were selected using univariate analysis, multivariate logistic regression, and LASSO. Nine machine-learning classifiers were trained and evaluated by AUC/accuracy, with model assessment using calibration, decision curve analysis (DCA), and SHAP. Calcium oxalate stones increased by 12.4% at recurrence, while infectious, uric acid, and other stones decreased by 5.5%, 5.9%, and 0.7% (P > 0.05). Infectious stones exhibited the highest transformation: only 3.2% remained unchanged, with 45.2% converting to calcium oxalate and 46.8% to uric acid stones. Compared with the changed group, the unchanged group was older and had a higher initial uric acid stone proportion and coronary heart disease prevalence; the changed group had a higher initial calcium oxalate proportion and more hypomagnesemia (P < 0.05). Multivariate analysis showed that age, initial stone composition, hypocitraturia, hypomagnesuria, and hypomagnesemia were independently associated with composition change, while hyperoxaluria and hyperlipidemia were inversely associated (all P < 0.05). LASSO retained seven core variables. Among the tested models, Random Forest showed the highest performance (training AUC 0.978, accuracy 0.979; test AUC 0.882, accuracy 0.847), with acceptable calibration; DCA suggested a net benefit > 0.6 across thresholds 0.2–0.6. SHAP ranked hyperoxaluria, initial composition, and age as the top contributors. In this cohort, recurrent stone composition change was associated with multiple metabolic and clinical factors, with hypocitraturia, hypomagnesuria, and hypomagnesemia showing independent associations. Among the evaluated algorithms, the Random Forest model achieved the highest performance metrics, and SHAP was used to describe the relative contribution of key variables. Given that this study is a single-center retrospective cohort with a limited sample size, the derived model may exhibit optimistic bias. Its robustness and generalizability therefore require further validation in larger-scale and external multi-center cohorts.