Abstract Sun507: Artificial Intelligence Driven Subphenotyping of In-Hospital Cardiac Arrest Patients Identifies Subgroups with Different Outcomes and Risk Factors: A Nation-Wide Analysis

Abstract

Introduction: Disparities in outcomes after in-hospital cardiac arrest (IHCA) across race, ethnicity, and socioeconomic status are impacted by social determinants of health (SDOH) and health care access. We applied unsupervised machine learning methodologies to identify novel patient-level IHCA subphenotypes by integrating clinical and sociodemographic data, and safety-net hospital (SNH) status. Methods: We utilized a cohort of 431,847 patients who suffered an IHCA from the American Heart Association Get with the Guidelines-Resuscitation Registry that included demographic characteristics, resuscitation outcomes and processes, and SDOH features. We implemented K-means clustering and hierarchical clustering with factor analysis for mixed data to identify subphenotypes. The primary outcome was survival to discharge and survival with good neurological status as defined by cerebral performance category (CPC) 1-2. We then compared survival outcomes across subphenotypes using Cox proportional hazards models. Results: Subphenotype A (n = 307,668) consisted of patients with high burden of comorbidities and the highest rates of unmet social needs. This subphrnotype had the lowest survival to discharge (19.1%) and good neurological outcomes proportion (12.2%) as compared to other subphenotypes (Figure 1). Subphenotype B (n = 57,356) was composed of patients who experienced high rates of respiratory disease and primarily arrested in the emergency department. Subphenotype C (n = 66,823) were more likely to undergo intubation during resusciation and demonstrated the best survival to discharge proportion (43.6%) and good neurological outcomes (32.2%). Subphenotype assignment was independently associated with survival (HR 0.766, 95% CI 0.755–0.776) and good neurological outcomes (HR 0.770, 95% CI 0.757–0.783) (Figure 2). Subphenotype A patients were more likely to receive care at safety-net hospitals (20.1%) compared to subphenotype C (17.1%). Treatment at a SNH was an independent predictor of lower survival (AUROC 0.82) and poorer neurological recovery (AUROC 0.80) (Figure 3). Conclusion: Using a large IHCA registry we identified three IHCA subphenotypes with significant differences in survival and neurological outcomes. By combining hospital structural features with detailed clinical and resuscitation data, this study offers a better understanding of IHCA risk across institutions and populations to inform targeted interventions.

Ähnliche Arbeiten