C22-14 Developing a Simplified Predictive Model for Post Cardiac Arrest Survivability to Personalize Code Status Conversations

Abstract Rationale Cardiopulmonary resuscitation (CPR) is the most consequential intervention used for cardiac arrest (CA), both in and out of the hospital, and is associated with a high risk of mortality. Code status decisions—the choice to accept CPR in the event of CA—are driven by patients’ personal preferences and values, with less attention paid to their personalized risk factors. These discussions are frequently not personalized to each patient’s post-CA mortality risk due to the heterogeneity of patient health status and difficulty estimating risk in real-time. We sought to develop a simple and reproducible model to predict patients’ 90-day post-CA mortality risk. Methods We used electronic health record data from a 11-hospital health system formatted using the Common Longitudinal ICU data Format (CLIF). We identified adults who sustained CA and received CPR using International Classification of Diseases, Tenth revision (ICD-10) and Current Procedural Terminology (CPT) codes from 2011 to 2025. We excluded patients who transitioned to do-not-resuscitate and comfort care. We also captured the pre-arrest severity of illness one day prior to the onset of arrest using the SOFA score, as well as the use of life-sustaining therapies (maximum norepinephrine equivalent dose, mechanical ventilation, and dialysis). We used a bidirectional stepwise logistic regression to create a 90-day post-CA risk model, trained and validated using a 70:30 split. Results There were a total of 593 patients identified. The median age was 66 IQR 55-74, 38.2% were female, 77.4% identified as non-Hispanic White, and 91% with English as their primary language. Prevalent Elixhauser-based comorbid conditions amongst the cohort included hypertension, fluid and electrolyte disorders, and anemia. The overall post-CA 90-day mortality was 70.4%. The model narrowed from an initial model with 37 covariates to 11 covariates including age, valvular disease, depression, and pre-arrest SOFA score. The statistically significant covariates included pre-arrest SOFA score with an odds ratio of 1.07 (95% CI: 1.01 - 1.15). When evaluated on the test set, the model’s AUROC was 0.59 (95% CI: 0.48 - 0.70). Conclusions We demonstrate the feasibility of a simple 90-day post-CA mortality prediction model. A similar model could potentially be incorporated into an electronic health record interface—specifically within the code status order set—to enhance providers’ a priori intuition about individual patients’ post-arrest mortality risk. Future studies should improve and enhance the predictive capacity using a federated data analytics approach to create an omnibus model. This abstract is funded by: None