EP665 - Predicting post-operative sepsis in IBD patients undergoing colectomies using machine learning: A NSQIP-IBD study

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.
• 

  Sonia L. Ramamoorthy, MD (she/her/hers)

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

  Disclosure information not submitted.
• 

  Nicole Lopez, MD

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

  Disclosure information not submitted.
• LP

  Lisa Parry, MD

  La Jolla, CA

  Disclosure information not submitted.
• 

  Shanglei Liu, MD (he/him/his)

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

  Disclosure information not submitted.
• BA

  Benjamin Abbadessa, MD

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

  Disclosure information not submitted.
• 

  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)

• 

  Samuel Eisenstein, MD

  Professor
  UC San Diego Health
  UC San Diego Health
  La Jolla, CA, United States

  Disclosure information not submitted.

Purpose/Background: Patients with inflammatory bowel disease (IBD) are at elevated risk for postoperative complications due to underlying frailty and immunosuppressive exposure. Postoperative sepsis remains a major source of morbidity, prolonged hospitalization and deleterious outcomes. Therefore, the ability to predict or stratify patients at increased risk of postoperative sepsis could provide a way to effectively allocate resources to better care for these patients and potentially prevent the downstream complications. We sought to evaluate the performance of machine learning algorithms in predicting postoperative sepsis and/or septic shock in IBD patients undergoing colectomies.

Methods/Interventions: Patients from the 2017-2023 NSQIP-IBD database who underwent a colectomy (CPT-based) for Crohn's disease or ulcerative colitis were included. Those with preexisting infections or sepsis were excluded. Datasets with < 30% missing data were imputed using multivariate imputation by chained equations. Over 50 demographic, clinical, surgical, and anesthetic variables were analyzed (124 features after one-hot encoding). Data were split 80:20 into training/test sets and modeled using logistic regression (LR), extreme gradient boosting (XGBoost), and neural networks. Model performance was evaluated by AUC, F1-score, precision, recall, and accuracy. SHAP (SHapley Additive exPlanations) analysis identified top predictive features.

Results/Outcomes: Among 4,256 patients (3,164 Crohn's, 1,092 UC), 124 (2.9%) developed postoperative sepsis or septic shock. Models were developed for two cohorts: Crohn's and UC. Logistic regression achieved the best discrimination (Crohn's: AUC = 0.775; UC: 0.850), outperforming XGBoost (0.771; 0.697) and neural networks (0.540; 0.735). After threshold optimization, the Crohn's LR model yielded accuracy = 0.814, sensitivity = 0.688, specificity = 0.817, precision = 0.089, and F1 = 0.157. Key features included alkaline phosphatase and operative duration (positive association) and sodium, total bilirubin, and biologic use within 60 days preoperatively (negative association). The UC LR model achieved accuracy = 0.772, sensitivity = 0.889, specificity = 0.767, precision = 0.140, and F1 = 0.242; ASA class, operative duration, and total bilirubin were most predictive of sepsis.

Conclusion/Discussion: We demonstrate the ability of our machine learning model to accurately predict postoperative sepsis in IBD colectomy patients despite significant class imbalance (2.9% positive rate), performing 3-5× better than chance alone. Baseline laboratory values, such as alkaline phosphatase, total bilirubin, and sodium remained strongly predictive values, while operative duration and ASA were positively associated with development of sepsis. Lastly, biologic use seems to be inversely associated with development of sepsis potentially demonstrating that better disease control may be associated with a decreased risk of postoperative sepsis. Lastly, our models' high sensitivity supports potential clinical utility for perioperative risk stratification in accurately identifying patients at increased risk and allocating appropriate resources. Prospective validation is warranted to assess real-world integration and generalizability to a larger cohort.

Figure 1: AUC-ROC for both Crohn's and UC colectomies, comparing logistic regression against xgboost and neural networks.
Figure 2: Top 5 model features for each best performing model (logistic regression) in each cohort (Crohn's and UC).