Using Machine Learning to Predict Risk Factors for 30-day Complication After Ureteroneocystostomy

Thursday, September 21, 2023

3:10 PM – 3:20 PM CT

Location: Avenue Ballroom, 4th Floor

Member Speaker(s)

Edward Wu, MS

OMS-4
Alabama College of Osteopathic Medicine

Disclosure(s): No financial relationships to disclose

Slides

Disclosure(s):

Edward Wu, MS: No financial relationships to disclose

Learning Objectives:

List potential complications following ureteroneocystostomy
Describe the concept of the random forest classification model
Identify risk factors for 30-day complication rate following ureteroneocystostomy in the pediatric population

Introduction/Purpose: Vesicoureteral reflux (VUR) is a retrograde flow of urine from the bladder to the kidneys. Ureteroneocystostomy (UNC) is considered by many to be the gold standard for surgical treatment of patients with VUR. Due to the small volume of UNC performed per year, the 30-day complication rates can be underpowered and underrepresented. Our study aims to predict risk factors for 30-day complication rates using random forest machine learning for pediatric cohort undergoing UNC for VUR. 

Methods or Case Description: The 2020 National Surgical Quality Improvement Program (NSQIP) Pediatric data was analyzed for patients undergoing UNC for VUR. Postoperative complications were analyzed and composite 30-day complication rates. Basic statistics were performed between the complication vs. no complication group. For categorical variables, Pearson chi square tests were performed, for continuous variables, mean+/-SD were presented, and ANOVA analysis was performed. A random forest model was used to predict the independent perioperative and preoperative variables. Variable of importance of reported using the random forest model that was ultimately included in the multivariate logistic regression model.

Outcomes: A total of 1,742 patients were evaluated. The mean age was 1630 days, 1155 (66.3%) were females, and 587 (33.7%) were males. The 30-day complication rate occurred in 275 (15.8%) of patients. Urinary tract infection (UTI) was the most common complication occurring at a rate of 3.8%. Other uncommon complication included superficial incisional surgical site infection (SSI) (0.29%), deep incisional SSI (0.06%), organ space SSI (0.23%), deep wound disruption (0.11%), renal insufficiency (0.34%), cardiac arrest (0.06%), bleeding requiring transfusion (0.40%), C.diff (0.17%), sepsis (0.63%), and blood stream infection (0.11%). After discharge, the postoperative emergency department rates were 9.41% with 4.25% of patients requiring readmission (Table 1). The unplanned reoperation rates were 1.84% and the mortality rate was 0.12% occurring in two patients. In the adjusted analysis, 35-36 weeks prematurity (OR 2.37, 95% CI 1.32-4.13) and ureteral stents removed in the operating room (OR 1.84, 95% CI 1.16-2.90) were associated with 30-day complications. Patient weights of >42 pounds were associated with decreased odds of compilation compared to 29-42 pounds (OR 0.46, 95% CI 0.25-0.84), and there was greater length of stay for those that had complications (OR 2.36, 95% CI 1.62-3.40) (Figure 1).

Conclusion:

Complication rate occurred in 15.8% of patients undergoing UNC for VUR. Machine learning using the Random Forest technique identified prematurity and stent removal in the operating room as risk factors for 30-day complication rate. Higher weight was protective against complications and those with complications showed longer length of hospital stay. We show that using machine learning can be an effective tool to predict postoperative outcomes in UNC.