(UroToday.com) The 2022 American Urological Association (AUA) Annual Meeting featured work from Mr. Miki Haifler and colleagues assessing the ability of a machine learning model to predict surgical intervention in symptomatic <5 mm ureteral stones. One of the most common reasons for emergency admission to the urology department is renal colic due to ureteral stone. Stone passage is heavily dependent on stone size and stone location. Ureteral stones have a 75%-89% expulsion rate and distal stones <5 mm have a 71%-98% rate of expulsion. According to the EAU guidelines, 95% of all ureteral stones <5 mm are expected to spontaneously pass within 40 days. On the other hand, the Journal of Endourology expects stones <5 mm to pass within 4 weeks and surgical intervention is only recommended after those 4 weeks. As seen, there are no definite guidelines for surgical intervention of ureteral stones <5 mm. As such, Mr. Miki Haifler and colleagues aim to identify the parameters that may predict the need for surgical intervention for ureteral stones <5 mm in attempt to create a machine learning model.
Patients from 2016 to 2021 with renal colic caused by ureteral stones were retrospectively reviewed. Data on demographic, imaging, laboratory and clinical history were collected. The cohort of 471 patients were included and randomly split into a training test (20%) and training set (80%). The training set was trained to predict the need for intervention using an extreme gradient boosting ML model (XGboost).
The performance of the model was assessed with the ROC curve metric (AUC) and decision curve analysis. Important variables were assessed with Shapley values additive explanation (SHAP).
Of 471 patients, 160 (34%) of patients underwent surgical intervention. Seventy-four percent of stones were in the distal ureter and the median stone diameter was 3.5 mm. There was no significant difference between the time to surgical intervention and the time to stone passage. The trained model had an AUC of 0.8 and 0.78 for the training and test data respectively.
The length of hospital stay and number of ER visits before admission were significantly different between those who underwent surgical intervention and those who did not.
The variables of highest importance based on SHAP were stone location and size.
Through this study investigating a machine learning model for the prediction of surgical intervention among patients with intractable pain caused by < 5 mm ureteral stones, Mr. Miki Haifler and colleagues presented the following conclusions:
- The prediction ML model was able to accurately predict stone passage and can aid in the decision making process of surgical intervention.
- Stone size, proximal stone location, heart rate, and number of ER visits contributes to the decision of surgical intervention.
After the presentation, the moderators questioned the purpose of machine learning. Currently, there are graphs that physicians can use to make the decision for surgical intervention. Hence, how does this machine learning model differ from what is currently available. Mr. Miki Haifler explains that the machine learning model differs in that it can capture relationships between different variables and utilize those relationships in its decision making.
The moderators ended the presentation stressing the importance of validation. Recently, machine learning has become popular in the field of medicine. There are numerous machine learning studies, but none that have been validated. External validation is an extremely important next step for the future of machine learning.
Presented by: Mr. Miki Haifler, Ramat Gan, Israel
Written by: Minh-Chau Vu, B.S., Department of Urology, University of California, Irvine @MinhChauVuuu on Twitter during the 2022 American Urological Association (AUA) Annual Meeting, New Orleans, LA, Fri, May 13 – Mon, May 16, 2022.