Introduction: Renal colic due to ureteral stone is among the most common reasons for acute admissions to a hospital urology department. A high expulsion rate is reported for ureteric stones =5 mm. We aim to fit a machine learning model (ML) for surgical intervention prediction in cases of intractable pain caused by <5 mm ureteral stones.
Methods: We retrospectively reviewed all patients with renal colic caused by ureteral stone <5 mm at our institute between 2016-2021. Data on age, sex, body mass index, the presence of associated hydronephrosis, fat stranding, stone side and location, medical history, serum white blood and neutrophil counts, creatinine, C-reactive protein, and vital signs were obtained upon admission.
Continuous and categorical variables were compared with the Wilcoxon rank-sum test and Fisher exact test. Missing data were imputed with multiple imputations.
We fit an extreme gradient boosting ML model (XGboost) to predict the need for intervention based on clinical parameters. Variable importance was assessed with Shapley values additive explanation. The performance of the XGboost was determined by the area under the ROC curve metric (AUC) and decision curve analysis.
Results: 471 patients (median age 49 [IQR 40.5 ,60], 83% males) were included. The median stone diameter was 3.5 mm (3-4.1) and 74% of the stones were in the distal ureter. 160 (34%) patients who experienced intractable pain underwent surgical intervention. The time to stone expulsion and the time to surgical intervention were not significantly different.
The model’s AUC according to the training and test data AUC were 0.8 and 0.78, respectively (p=0.68, Figure 1). Stone location and size had the highest variable importance compared to other variables.
Conclusions: The decision to surgically remove ureteral stones < 5 mm is difficult and can be facilitated by the use of our accurate prediction ML model. Our prediction ML model was both accurate and generalizable. The most important variables for intervention prediction were stone size and location. The presented model may be implemented in the treatment protocols, helping the decision-making process.
.jpg)
.jpg)
