Introduction: There are limited predictive markers of stone recurrence, which poses a challenge for the clinical management of stone disease. Furthermore, inability to predict stone events is a significant limitation for clinical trials, where many patients must be enrolled to obtain sufficient stone events for analysis. In this study, we sought to use machine learning methods to identify a novel algorithm to predict stone recurrence. Methods: Patients enrolled in the Registry for Stones of the Kidney and Ureter (ReSKU), a multi-institutional registry of nephrolithiasis patients collected between 2015-2020, with at least one prospectively collected 24-hour urine test (Litholink) were included in the training set. A test set of stone patients not enrolled in ReSKU with 24-hour urine test results were collected. Stone events were defined as either an office visit where a patient reported symptomatic passage of stone(s) or surgery for stone removal. Seven prediction classification methods were evaluated: Decision Tree, Logistic regression with ElasticNet, Extra Trees, kNeighbors, LightGBM, Logistic Regression, and Random Forest. Probability of Recurrent Nephrolithiasis (PRN) score ranging from 0-1 was generated from the models. Predictive analyses and ROC curve generation was performed in R. Results: A training set of 423 kidney stone patients with stone event data and 24-hour urine samples were trained using the prediction classification methods. The highest performing prediction model was a Logistic Regression with ElasticNet machine learning model (AUC = 0.65; Figure 1A). Selecting for high confidence predictions (PRN 0 - 0.15, 0.8 - 1.0) improved model accuracy (AUC = 0.82; Figure 1B). The model was validated on a test set of 172 stone patients. Patients predicted to have high likelihood of stone recurrence had an increase in the average number of stone events compared to those with low likelihood (3.6 vs 1.7 events; p = 2e-7). Prediction accuracy in the test set demonstrated moderate discriminative ability (AUC = 0.64; Figure 1C). Conclusions: Using machine-learning trained on 24-hour urine and stone outcome data, a novel algorithm was developed to predict stone recurrence. Initial validation studies suggest ability of algorithm to identify patients with higher likelihood of stone events and warrants further study. SOURCE OF Funding: n/a
.jpg "Kevin Shee, MD, PHD photo")
