Comparative Evaluation of Machine Learning Models for Reliable Kidney Stone Prediction: Performance, Robustness, and Clinical Utility

Abstract

Kidney stone disease nephrolithiasis represents a growing global health concern, necessitating accurate and early prediction to prevent complications and optimize healthcare resources. This study conducts a comprehensive comparative assessment of multiple supervised machine learning algorithms Logistic Regression, Decision Trees, Random Forests, Support Vector Machines, Gradient Boosting, and Neural Networks for predicting kidney stone formation. Model performance was systematically evaluated using a suite of statistical and clinical metrics to determine predictive accuracy, robustness, and interpretability. The findings indicate that ensemble-based models, particularly Random Forest and Gradient Boosting, consistently outperform other algorithms in terms of accuracy and generalization capability. In contrast, simpler models such as Logistic Regression and Decision Trees offer enhanced interpretability, making them more suitable for clinical decision support. To enhance transparency and clinical trust, SHAP (SHapley Additive exPlanations) analysis was employed to elucidate feature contributions, identifying serum calcium and uric acid as the most influential biomarkers in prediction outcomes. The study underscores the critical balance between predictive performance, model robustness, and interpretability, demonstrating that explainable and validated machine learning frameworks can serve as effective tools for early risk stratification, timely clinical intervention, and improved patient management in nephrolithiasis care.