Abstract 265: Social Determinants Of Health Impact Phenotype-based Outcome Prediction After Pediatric Arrest

Abstract

The Personalizing Outcomes after Child Cardiac Arrest-1 (POCCA-1) trial showed that blood-based brain injury biomarkers have promise to prognosticate outcome post-pediatric cardiac arrest. However, whether POCCA-1 biomarkers perform accurately among all children is unknown. Hypothesis We hypothesized that POCCA-1 biomarkers will have different performance accuracy by distinct clinical phenotypes after pediatric cardiac arrest. Methods: Fourteen US centers enrolled 164 children aged < 18 y with pediatric intensive care unit admission post-arrest. Glial fibrillary acidic protein (GFAP), ubiquitin carboxyl-terminal esterase L1, neurofilament light, and Tau proteins were measured in blood post-arrest d1-3. Patient ZIP codes linked to US census data provided socioeconomic neighborhood characteristics. We analyzed 120 children with evaluable data (n=50, unfavorable outcome [Vineland Adaptive Behavioral Scale < 70 or dead]) to create 3 clinical phenotypes using latent class analysis. POCCA-1 biomarker levels and outcomes were analyzed by phenotype. The POCCA-1 model was analyzed to predict 1 y unfavorable outcome using multivariate AUROC by phenotype for all 4 biomarkers on d1- 3 post-arrest. Results: Group 1 (n=35, 22%) were older, of White race, with private insurance and out-of-hospital cardiac arrests (OOHCA), fewer fever post-arrest, living in neighborhoods of mostly White race and lower poverty. Group 2 patients (n=77, 47%) were younger, of White race, with in-hospital cardiac arrests living in more socioeconomically diverse neighborhoods. Group 3 (n=51, 31%) were diverse in race and age, of more male sex, with Medicaid insurance and asphyxial, OOHCA, living in more impoverished neighborhoods. Unfavorable outcome was higher in Group 3 (69%) vs 1 (48%) and 2 (28%), p<0.001. All biomarkers (except Tau) were significantly increased on d1-3 in Group 3. Biomarker accuracy to predict outcome was excellent for Group 1 with 11/12 AUROCs significant. Models were less accurate for Group 2 (7/12) and marginal for Group 3 (3/12). Conclusions: The POCCA-1 prediction model performance varied by patient, arrest, and neighborhood context, suggesting that prediction models need to be personalized using clinically relevant and equitable phenotypes.