EP795 - Predicting Clostridium Difficile infection in post-operative IBD patients using machine learning: A NSQIP-IBD study

Monday, May 12, 2025

1:53 PM – 1:58 PM PT

ePoster Presenter(s)

NN

Nicholas C. Neel, MD

General Surgery Resident
UCSD Department of Surgery
San Dimas, California, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.

Author(s)

BP

Brian Park, MD

Assistant Physician
UCSD Department of Anesthesia
San Diego, California, United States

Disclosure information not submitted.
Nicole Lopez, MD

Associate Professor
UCSD Department of Surgery
San Diego, California, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.
LP

Lisa Parry, MD

La Jolla, CA

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.
Sonia Ramamoorthy, MD (she/her/hers)

Professor
UCSD
La Jolla, California, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no relevant relationships with ineligible companies.
BA

Benjamin Abbadessa, MD

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.
Shanglei Liu, MD (he/him/his)

Assistant Professor
UCSD Department of Survery
San Diego, California, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.
Bard C. Cosman, MD, MPH

Professor of Clinical Surgery
University of California San Diego
San Diego, CA, United States

All of the relevant financial relationships listed below have been mitigated.
This individual has no financial relationships with ineligible companies.
RG

Rodney Gabriel, MD, MAS

Associate Professor
UCSD Department of Anesthesiology
San Diego, California, United States

Disclosure information not submitted.

Senior Author(s)

SE

Samuel Eisenstein, MD

Associate Professor
UC San Diego Health
La Jolla, California, United States

All of the relevant financial relationships listed below have been mitigated.
Takeda: Consultant (Ongoing)

All of the relevant financial relationships listed below have been mitigated.

Nicholas C. Neel, MD: This individual has no financial relationships with ineligible companies.

Nicole Lopez, MD: This individual has no financial relationships with ineligible companies.

Lisa Parry, MD: This individual has no financial relationships with ineligible companies.

Sonia Ramamoorthy, MD: This individual has no relevant relationships with ineligible companies.

Benjamin Abbadessa, MD: This individual has no financial relationships with ineligible companies.

Shanglei Liu, MD: This individual has no financial relationships with ineligible companies.

Bard C. Cosman, MD, MPH: This individual has no financial relationships with ineligible companies.

Samuel Eisenstein, MD: Takeda: Consultant (Ongoing)

Image or Table (Do Not add text here):

Purpose/Background: IBD patients are at increased risk for post-operative complications given their preoperative frailty and exposure to immune suppressants. Clostridium difficile infection (CDI) has been shown to have a higher prevalence in post-operative IBD patients with increased morbidity, length of stay, and worse overall outcomes. Previous studies have tried to identify risk factors for CDI infection but these are largely single center and limited in scope. We examined the ability of machine learning algorithms to correctly identify patients at risk for developing post-operative CDI.

Methods/Interventions: All patients from the NSQIP-IBD 2017-2022 cohort were included in analysis with the primary outcome of interest being CDI within 30 days of surgery. Patients with local or systemic infections prior to surgery were excluded from the study. Over 50 demographic, medical, surgical, and anesthetic characteristics, were included (124 features after expanding to binary features). The data was split into 90:10 training:test set using three different machine learning models (logistic regression, random forest, and neural networks) to predict CDI as the outcome of interest. Each model was assessed by area under the curve (AUC) score, F1 score, precision, recall, and accuracy (Table 1). SHAP (SHapley Additive exPlanation) charts were used to assess important model features and analyze their clinical relevance. Subgroup analysis was performed for patients undergoing subtotal colectomy and ileocolic resection.

Results/Outcomes: Of the 6,879 eligible patients, 6,182 patients were included for analysis, of which 443 (7.2%) were diagnosed with CDI postoperatively. In the entire cohort, the neural network algorithm had the best prediction with AUC score of 0.78 and F1 score of 0.32 (Table 1). Top patient characteristics of the model include sex, ulcerative colitis, and white blood cell count. Top surgical features include proctectomy procedure, duration of surgery, and laparoscopic approach, while the use of oral antibiotic prep was positively correlated with prediction of CDI. Subgroup analysis found 1,709 patients that underwent subtotal colectomy of whom 98 (9.1%) developed CDI, with logistic regression showing the best performance with an F1 score of 0.44 and AUC of 0.81 (Table 1). Top features based on SHAP analysis include white blood cell count, oral antibiotic prep, sex, hematocrit, and albumin. Lastly, 2,054 patients underwent ileocolic resection with 154 (7.5%) developing CDI. Logistic regression performed the best in this cohort with F1 score of 0.53, AUC 0.92 (Table 1). Top features based on SHAP analysis include white blood cell count, sex, antibiotic prep, and wound class.

Conclusion/Discussion: We demonstrate the employment of machine learning to accurately predict post-operative CDI in IBD patients with improved accuracy when surgical procedures are examined individually. Model features show some expected values (such as WBC, albumin, length of procedure), while other factors may demonstrate interesting correlations including that oral antibiotic prep with or without mechanical bowel prep significantly increases the risk of postoperative CDI in this population. Further study with prospectively collected data is needed to validate these algorithms.