Poster 050: Development and Validation of a Deep Learning–Enhanced Survival Prediction Model for Critically Ill Patients Using EHR-Derived Vitals, Labs, and Comorbidities

Background: Survival prediction in critical-care currently utilizes static scoring systems like APACHE II and SOFA. This fails to capture the evolution of critically-ill patients in dynamic conditions like shock. We aimed to develop and validate a deep-learning model to predict 6-month survival by integrating longitudinal time-series data with static clinical features.

Methods: We conducted a retrospective cohort study of 4,325 adult ICU patients across 13 hospital sites in the OMOP CDM v2 database. Our approach utilized a three-stage ensemble architecture with LSTM for time-series data, DeepSurv for static features, and logistic regression for calibrated survival probabilities. The primary outcome was 6 month survival. SHAP was used for model interpretability.

Outcome: The model achieved AUROC of 0.9133, accuracy of 90.4%, sensitivity of 96.2%, and F1 score of 0.9423. SHAP analysis identified VTE, oxygen saturation, diabetes, sepsis, and AKI as the most influential predictors, with distinct patterns across glycemic subgroups. This highlighted complex interplay between glucose regulation and organ failure.

Conclusion: A deep learning–enhanced ensemble model used to synthesize temporal and static EHR data outperformed traditional, static ICU scoring systems. This framework offers a scalable foundation for real time decision support.