Abstract C025: Integrating health disparities and AI/ML-driven survival prediction in critically ill patients with multiple primary malignancies

Abstract Patients with multiple primary malignancies (MPMs) admitted to the intensive care unit (ICU) represent a clinically complex and understudied subgroup within oncology, often presenting with overlapping treatment histories, comorbidities, and heightened risk of critical care interventions. Prior studies have demonstrated that Black and Hispanic patients with multiple primary cancers experience higher risks of cancer-related and cardiovascular mortality compared to White patients, likely due to differences in access to care, comorbid burden, and stage at presentation. In a cohort of 1,249 ICU patients with MPMs, the number of distinct cancers per patient ranged from 2 to 13, with the most common malignancies involving the brain, respiratory tract, pituitary gland, and central nervous system. Kaplan–Meier survival analysis revealed a steep decline in survival probability following ICU admission, although a subset of patients achieved more stable long-term outcomes. To address the heterogeneity and clinical complexity of this population, we applied multiple machine learning (ML) models—including Cox proportional hazards, XGBoost, and DeepHit—to predict 28-day survival and identify key prognostic factors. XGBoost demonstrated exceptional discriminative performance (AUC = 0.989, Brier score = 0.127), while DeepHit showed limited predictive value in preliminary testing. To enhance interpretability, SHAP (SHapley Additive exPlanations) values were used to evaluate both individual and global feature contributions to survival prediction. Hyperglycemia within the first 24 hours of ICU admission was significantly associated with increased mortality risk (HR = 1.16, p = 0.03), while insulin administration and comorbidities such as cancer, malnutrition, and sepsis emerged as strong predictors of poor outcomes. Time-varying covariate analysis revealed dynamic risk profiles, with cancer and malnutrition showing increasing hazard over time, and sepsis exerting an immediate and sustained adverse effect. These findings underscore the limitations of static models and the need for real-time, individualized risk assessment in this high-risk population. Importantly, our study demonstrates that advanced ML models, particularly XGBoost combined with SHAP explanations, can provide transparent, patient-specific survival predictions regardless of racial background. By identifying the most influential survival predictors for each patient, these tools enable clinicians to deliver personalized, equitable critical care and help overcome previously observed racial health disparities in ICU outcomes for patients with multiple primary malignancies. Citation Format: Joab O. Odera. Integrating health disparities and AI/ML-driven survival prediction in critically ill patients with multiple primary malignancies [abstract]. In: Proceedings of the 18th AACR Conference on the Science of Cancer Health Disparities; 2025 Sep 18-21; Baltimore, MD. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2025;34(9 Suppl):Abstract nr C025.