SSAT - MACHINE LEARNING PREDICTION OF INCREASED HEALTHCARE RESOURCE UTILIZATION IN COLECTOMY FOR INFLAMMATORY BOWEL DISEASE.

Back to 2025 Abstracts

MACHINE LEARNING PREDICTION OF INCREASED HEALTHCARE RESOURCE UTILIZATION IN COLECTOMY FOR INFLAMMATORY BOWEL DISEASE.
Gustavo Salgado-Garza^*, Cynthia Araradian, Maura Walsh, Shahrose Rahman, Elliot Ballato, Sandy Fang, Brett C. Sheppard, Patrick J. Worth, Vassiliki L. Tsikitis
Department of Surgery, Oregon Health & Science University, Portland, OR

Background
Healthcare spending for patients with inflammatory bowel disease (IBD) has steadily increased over the past few decades. While some of the increase is due to the rise in the use of biologics - surgery and postoperative care contribute significantly to healthcare spending. We aimed to analyze the preoperative factors associated with increased healthcare resource utilization (HRU). To accomplish this, we developed and internally validated machine learning prediction models for increased HRU.
Methods
We analyzed the ACS-NSQIP database from 2018 to 2022 to identify preoperative factors that could predict increased HRU in patients undergoing non-emergent colectomy for Crohn�s disease or ulcerative colitis. Increased HRU was defined as a composite of either a) discharge destination other than home, b) readmission within 30 days of surgery, or c) prolonged length of stay. A total of six machine learning algorithms were utilized to find the best-performing model determined by accuracy. Models were internally validated with an 80:20 training and testing split.
Results
A total of 7,535 patients were analyzed. The overall rate of increased HRU was 33%. The best performing model was the random forest, achieving an AUC of 0.75, an accuracy of 0.73, and a F1 of 0.81 (Figure 1). The most important preoperative variables for model accuracy were hematocrit, albumin, and blood urea nitrogen (Figure 2).
Conclusion
Machine learning algorithms, particularly random forest models, can accurately predict increased healthcare resource utilization in patients undergoing colectomy for IBD. Implementing of our predictive model in the preoperative setting could enhance patient education and improve healthcare system preparedness before surgery.

Figure 1. Receiver operating characteristic curve for the final random forest model. AUC = area under the curve.

Figure 2. Importance of predictor variables for model accuracy. The higher the mean decrease in accuracy, the more important the variable is for the overall random forest model.

Back to 2025 Abstracts