Incorporating Imaging Markers of Brain Health in Modeling of Functional Outcome After Acute Ischemic Stroke: A Quantitative Comparison Study

Background and Objectives Brain health facilitates resilience to withstand detrimental effects of brain disease, but is challenging to assess in the time-sensitive setting of acute ischemic stroke (AIS). In this international observational multicenter study, we compare quantitative MRI markers of structural brain health in their ability to predict functional outcome after AIS. Methods We included AIS patients from the international MRI-GENIE study (multicenter; 2003-2011) with acute T2-FLAIR imaging. Automated pipelines estimated white matter hyperintensity volume (WMHv), brain volume, and intracranial volume (ICV). We normalized WMHv by individual brain volume, creating WMH load for each patient. Assessed brain health markers included: brain parenchymal fraction (brain volume relative to ICV); radiomics derived brain age (based on an ElasticNet linear regression model); brain reserve (normal appearing brain volume [brain volume minus WMHv] relative to ICV), and effective Reserve (eR, latent variable based on age, WMH load [WMHv and brain volume). We added the markers to a clinical reference model (including age, sex, diabetes type 2, hypertension, NIHSS, atrial fibrillation and prior stroke), comparing model performances between separate multivariable regression models in their prediction of unfavorable outcome (90-day modified Rankin Scale score 3-6), using Bayesian Information Criterion (BIC). Results We analyzed 2,303 patients (median age 66 years, 46% female, 27% unfavorable outcome) after excluding 357/2,660 patients (13.4%) due to missing follow-up data. Comparison using BIC provided strong statistical evidence (ΔBIC > 6) for all four models using a brain health marker to outperform the clinical reference model (BIC=2354.9). The eR model showed the lowest BIC value (BIC=2318.4), providing very strong (ΔBIC > 10) statistical evidence to outperform all other models. The model utilizing radiomics derived brain age showed the second lowest BIC value (BIC=2329.3), providing very strong (ΔBIC > 10) statistical evidence to outperform all other models except the eR model. Discussion Incorporating quantitative MRI markers of brain health in predictive models enhance personalized outcome prognostication after AIS. Our results suggest eR as a viable marker for future studies investigating structural brain health and ischemic stroke.