Unsupervised learning of early post-arrest brain injury phenotypes

Abstract

Introduction: Trials may be neutral when they do not appropriately target the experimental intervention. We speculated multimodality assessment of early hypoxic-ischemic brain injury would identify phenotypes likely to benefit from therapeutic interventions.

Methods: We performed a retrospective study including comatose patients resuscitated from out-of-hospital cardiac arrest (OHCA) by one of 126 emergency medical services or in-hospital arrest at one of 26 hospitals from 2011 to 2019. All patients were ultimately transported to a single tertiary center for care including standardized initial neurological examination, brain imaging and electroencephalography; targeted temperature management (TTM); hemodynamic optimization targeting mean arterial pressure (MAP) >80 mmHg; and, coronary angiography for clinical suspicion for acute coronary syndrome. We used unsupervised learning to identify brain injury phenotypes defined by admission neurodiagnostics. We tested for interactions between phenotype and TTM, hemodynamic management and cardiac catheterization in models predicting recovery.

Results: We included 1086 patients with mean (SD) age 58 (17) years of whom 955 (88%) were resuscitated from OHCA. Survival to hospital discharge was 27%, and 248 (23%) were discharged with Cerebral Performance Category (CPC) 1-3. We identified 5 clusters defining distinct brain injury phenotypes, each comprising 14% to 30% of the cohort with discharge CPC 1-3 in 59% to <1%. We found significant interactions between cluster and TTM strategy (P = 0.01), MAP (P < 0.001) and coronary angiography (P = 0.04) in models predicting outcomes.

Conclusions: We identified patterns of early hypoxic-ischemic injury based on multiple diagnostic modalities that predict responsiveness to several therapeutic interventions recently tested in neutral clinical trials.

Keywords: Cardiac arrest; Clustering; Outcomes; Phenotype; Precision medicine; Unsupervised learning.