P0287 - Explainable Machine Learning Identifies Predictors of Mortality and Colectomy in Colon Ischemia Using SHAPley Additive Explanations (SHAP) and Xtreme Gradient Boosting (XGBoost)

Award: ACG Presidential Poster Award

Shanti Patel, MD¹, Lawrence J. Brandt, MD, MACG², Akhil Vaid, MD³, Ali Soroush, MD, MS³, Kanika Sehgal, MBBS¹, Ikechukwu Elvis Eze, MD⁴, Abdelkader Chaar, MD⁵, Rabia Rizwan, MD⁶, Savio John, MD, FACG⁵, Thiruvengadam Muniraj, MD⁴, Paul Feuerstadt, MD, FACG⁷
1Yale New Haven Health, New Haven, CT; 2Montefiore Medical Center, Bronx, NY; 3Icahn School of Medicine at Mount Sinai, New York, NY; 4Yale University School of Medicine, New Haven, CT; 5SUNY Upstate Medical University, Syracuse, NY; 6Houston Medical Center, Houston, TX; 7Yale University School of Medicine and PACT-Gastroenterology Center, Hamden, CT

Introduction: Colon ischemia (CI) is a common disorder that may necessitate colectomy or lead to mortality. Our objective was to establish a reliable prediction model to inform clinical decision-making and identify variables associated with 30-day colectomy or mortality. 

Methods: A multicenter database (Yale-New Haven Hosp, Montefiore Med Center, SUNY-Upstate) of inpatients with biopsy-proven CI was used to develop machine learning models predicting 30-day colectomy or mortality. Dataset included manually collected demographic, lab, imaging, and pathology variables at the time of presentation of patients diagnosed with CI. Five models (XGBoost, Random Forest, Logistic Regression, SVM, Neural Network) were evaluated. The best overall model, XGBoost, was retrained on the full dataset using default parameters. The top 20 predictors of 30-day mortality and colectomy were identified based on mean absolute SHAP values and visualized with beeswarm plots.

Results: Among 1,221 patients with CI, 30-day colectomy and mortality rates were 14.7% and 8.3% respectively. XGBoost demonstrated consistently strong and balanced performance across both 30-day colectomy and mortality predictions, performing best in predicting mortality. (Figure 1). SHAP analysis revealed distinct predictors for 30-day mortality and colectomy in patients with CI (Figures 2a and 2b). 
The most significant predictors of 30-day colectomy were male sex, fever ≥ 38.5 °C and ICU need. Key lab predictors were BUN > 20 mg/dL, serum sodium < 136 mEq/L, WBC > 15,000/μL, Hgb < 12 g/dL, and low serum albumin (g/dL). Additional predictors included mucosal ulceration (colonoscopy), necrosis (pathology), and left-sided colon disease (CT). For 30-day mortality, the best predictors were older age and need for ICU. Important lab predictors were WBC > 15,000/μL, BUN > 20 mg/dL, elevated RDW (%), Hgb < 12 g/dL, serum sodium < 136 mEq/L, reduced GFR (mL/min/1.73 m²), elevated lactic acid (mmol/L), prolonged INR, and beta blocker use.

Discussion: Our analysis showed XGBoost is a reliable and a well-balanced model to predict short-term outcomes in CI. 30-day colectomy was more strongly associated with endoscopic findings of severe mucosal injury and distribution of disease on CT, while 30-day mortality was linked to systemic organ failure, e.g., critical care need, renal impairment, metabolic derangement, and coagulopathy. Attention to these variables can potentially guide clinicians in managing higher -risk patients more aggressively to possibly improve outcomes.


Figure: Figure 1. Receiver operating characteristic (ROC) curves comparing five machine learning models for predicting (left) 30-day colectomy and (right) 30-day mortality in patients with colon ischemia.


Figure: Figure 2a and 2b. SHAP summary plots showing the top predictors of (a) 30-day colectomy and (b) 30-day mortality in patients with colon ischemia,

Disclosures:
Shanti Patel indicated no relevant financial relationships.
Lawrence Brandt indicated no relevant financial relationships.
Akhil Vaid indicated no relevant financial relationships.
Ali Soroush indicated no relevant financial relationships.
Kanika Sehgal indicated no relevant financial relationships.
Ikechukwu Elvis Eze indicated no relevant financial relationships.
Abdelkader Chaar indicated no relevant financial relationships.
Rabia Rizwan indicated no relevant financial relationships.
Savio John indicated no relevant financial relationships.
Thiruvengadam Muniraj indicated no relevant financial relationships.
Paul Feuerstadt: Ferring Pharmaceuticals – Advisor or Review Panel Member, Honoraria. Rebiotix, Inc. – Consultant. Seres Therapeutics – Advisor or Review Panel Member, Honoraria. Takeda Pharmaceuticals – Advisor or Review Panel Member, Consultant, Honoraria.

Shanti Patel, MD¹, Lawrence J. Brandt, MD, MACG², Akhil Vaid, MD³, Ali Soroush, MD, MS³, Kanika Sehgal, MBBS¹, Ikechukwu Elvis Eze, MD⁴, Abdelkader Chaar, MD⁵, Rabia Rizwan, MD⁶, Savio John, MD, FACG⁵, Thiruvengadam Muniraj, MD⁴, Paul Feuerstadt, MD, FACG⁷. P0287 - Explainable Machine Learning Identifies Predictors of Mortality and Colectomy in Colon Ischemia Using SHAPley Additive Explanations (SHAP) and Xtreme Gradient Boosting (XGBoost), ACG 2025 Annual Scientific Meeting Abstracts. Phoenix, AZ: American College of Gastroenterology.