Machine learning-based prediction of ICU mortality in patients with pancreatic cancer: A MIMIC-IV retrospective cohort study.

142

Background: Pancreatic cancer patients requiring ICU admission face high short-term mortality, yet no disease-specific prediction tools exist. General severity scores like SOFA may not capture pancreatic cancer-specific risk factors including biliary sepsis and malnutrition. We aimed to develop and validate ML models to predict ICU mortality in pancreatic cancer patients using MIMIC-IV. Methods: This retrospective study used MIMIC-IV (v3.1, 2008-2022). Adults with pancreatic cancer (ICD-9: 157.x; ICD-10: C25.x) and ICU stays ≥6 hours were included; neuroendocrine tumors were excluded. Features included demographics, tumor characteristics, comorbidities, SOFA score, disease-specific complications, labs, and vitals from the first 24 hours. Primary outcome was ICU mortality. Four models (logistic regression, LASSO, random forest, XGBoost) were trained (2008-2016) and temporally validated (2017-2022). Performance was compared against SOFA alone. SHAP analysis identified key predictors. Results: 736 ICU stays from 591 patients were analyzed (training: 519; validation: 217). Mean age was 68.0±11.4 years; 57.5% male. Tumors were predominantly pancreatic head (63.0%) and 60.6% had metastatic disease. Biliary sepsis was the leading admission reason (38.3%). ICU mortality was 12.6%, hospital mortality 22.8%, and 28-day mortality 36.3%. All ML models outperformed SOFA alone (Table 1). Logistic regression achieved the highest AUROC (0.856) while random forest had the best calibration (Brier 0.084). SHAP analysis identified SOFA score, mean arterial pressure, SpO2, lactate, and respiratory rate as top predictors, with biliary obstruction and albumin among key disease-specific contributors. Risk stratification yielded a low-risk group (82.9%, 6.1% mortality) versus high-risk group (12.0%, 46.2% mortality). Conclusions: ML models incorporating pancreatic cancer-specific features significantly outperformed the SOFA score for predicting ICU mortality. Biliary sepsis dominated ICU admissions, and disease-specific factors including biliary obstruction and hypoalbuminemia contributed meaningfully to prediction. These models may support clinical decision-making regarding ICU triage and goals-of-care discussions in this high-risk population.