Abstract Sat306: External Validation of Out-of-Hospital Cardiac Arrest Outcome Prediction Models

Abstract

Background: Inappropriate early neuroprognostication of out-of-hospital cardiac arrest (OHCA) patients may lead to inappropriate early withdrawal of life-sustaining treatments. Several prediction models using early clinical data have been published, with varying degrees of external validation. Hypothesis: Out of six commonly used clinical predictive models, at least one may have an acceptable specificity and negative predictive value and can be used to identify patients who will have poor neurological outcome and mortality. Aims: The aim of this study was to perform an external validation of six different clinical predictive models—the Cardiac Arrest Hospital Prognosis score (CAHP), Swedish Cardiac Arrest Risk Score, Targeted Temperature Management score, Out of Hospital Cardiac Arrest score, Miracle2 score, and the Cardiac Arrest Survival Score—to determine which scores accurately predict neurological outcome and survival. Methods: Data were collected on OHCA patients who were part of the International Cardiac Arrest Registry (INTCAR) and received treatment at Maine Medical Center between 2021 and 2023. Data from INTCAR and the electronic medical record were used to calculate predicted outcomes with specificity and negative predictive value used to determine if the models were externally valid. Statistical software was used to recreate the formulas for each model and Delong’s test was also performed in order to compare the discriminatory ability between models. Results: 222 patients were included with 91 (41%) having a shockable initial heart rhythm and a mean time to return of spontaneous circulation of 28.73 minutes (± 16.48); 75 (34%) patients survived to hospital discharge. Variability in discrimination was observed with AUC values ranging from 67.3 to 77.4. Among the models studied, the CAHP score resulted in the highest specificity (82.3), indicating an increased ability to correctly identify individuals with poor outcomes. This model was less likely to incorrectly categorize poor outcomes, with a false negative value of 51.4%. The OHCA model had the highest AUC value (77.4% [CI 70.6%-84.2%]). The Miracle2 score exhibited the best discrimination, allowing it to maintain greater consistency across all thresholds. Conclusions: While some prediction models have undergone external validation to determine their accuracy, the results from the patient population used in this study indicate that all models may lead to errors when predicting patient outcome.