Aims: Preventive counseling for kidney stone disease requires accurate risk identification, yet traditional interpretation fails to capture complex metabolic interactions. This study aimed to develop a machine learning-based screening tool using routine blood and urine data to identify high-risk individuals and provide actionable insights for patient education.
Instrument & Methods: Data from 952 individuals (714 stone formers, 238 controls) with 28 features were analyzed. Particle Swarm Optimization selected features, balancing accuracy and test reduction. An XGBoost model was developed with 5-fold cross-validation. SHAP analysis identified influential predictors and their clinical implications.
Findings: The PSO-optimized model reduced tests from 28 to 8 features (71% reduction) with 93.49% accuracy (95% CI: 91.8-95.0%). The final panel included: Age, Blood_Phosphorus, Blood_Uric_Acid, Blood_Creatinine, eGFR, Urine_pH, Urine_Calcium_24h, and Urine_Citrate_24h. SHAP identified modifiable risk factors: low citrate (potassium citrate), elevated calcium (dietary modification), elevated uric acid (purine reduction), and reduced eGFR (renal evaluation).
Conclusion: This tool enables non-invasive risk identification from routine data, empowering targeted preventive counseling in primary care.