A Comprehensive Machine Learning Model for Kidney Stone Prediction

Kidney stone disease affects nearly 1 in 11 Americans over their lifetime and results in >$10 Billion of healthcare costs annually. In this paper, we propose machine learning models for the prediction of kidney stone presence in urine using five parameters of single urine samples: specific gravity, pH, osmolality, conductivity and concentrations of urea and calcium. Two types of features were engineered that capture the complex interaction between urine components (Specific gravity-to-calcium ratio and Calcium-to-pH product) and their relationship with kidney stone formation. We test the models with Logistic Regression, Support Vector Classifier(SVC), Random Forest and achieve the score of 70.65%, 72.83% and 71.74%. Based on precision-recall metrics, further evaluation indicated that SVC performed the best among other models that was highest in both stands at 0.78 for precision and 0.64 for recall. This presents a demonstration of the effectiveness and importance of machine learning models in medical diagnosis, in this case useful for non-invasive prediction of kidney stones for possible prevention or treatment.