Global Brain Health Predictors of Post-Stroke Sensorimotor Recovery using AI-Enhanced Clinical MRIs - PROJECT SUMMARY The goal of the proposed work is to improve clinical predictions of stroke recovery and, subsequently, precision rehabilitation, by developing powerful statistical models that use routine, baseline clinical assessments (clinical MRIs, demographics, and behavior) to predict motor and cognitive stroke outcomes at 3, 6, and 12 months. This work builds on results from our current R01 grant (2020-2025) which established the role of global brain health (GBH; measured across cellular, vascular, and glymphatic domains as measured by brain age, white matter hyperintensities [WMH], and enlarged perivascular spaces [PVS], respectively) in accurately predicting stroke recovery. Specifically, using longitudinal data from baseline to 3-months post-stroke, we find that poorer GBH at baseline predicts worse lesion damage and more severe sensorimotor impairment at 3 months, and, conversely, more severe lesion damage at baseline predicts worsened GBH at 3 months. Relatedly, using a large, cross- sectional database of patients from acute to chronic stroke, we confirm that GBH is strongly associated with sensorimotor outcomes, with strongest relationships in the chronic stage, after secondary atrophy has time to evolve. We also find that GBH mediates relationships between lesion damage and stroke outcomes, suggesting that interventions to improve brain health could potentially mitigate the impact of stroke damage. However, while GBH appears to be a clinically useful biomarker, it is challenging to translate these findings into the clinic because the study of GBH requires high-resolution MRIs which are typically only acquired as part of research studies. Therefore, in this study we aim to leverage recent advances in generative AI that will allow us to extract GBH metrics from routinely acquired clinical MRIs. This study has three specific aims. Aim 1 will optimize AI algorithms to generate synthetic, high-resolution metrics (MRIAI) from routine clinical MRIs (MRIc) specifically for people with stroke, utilizing data from 437 people with stroke who have both MRIc and research MRIs (MRIR) from the ENIGMA Stroke Recovery database. Aim 2 will define longitudinal trajectories of MRIAI and their associations with impairment over the first year of stroke, with the hypothesis that declines in functional outcomes between 6 to 12 months post-stroke are correlated with worsened brain health during that same time period. For this aim, we will recruit 210 people with stroke for baseline, 3-, 6-, and 12-month clinical brain scans with brief behavioral assessments. Aim 3 will develop clinical decision trees that utilize baseline AI-enhanced clinical MRIs to predict mild, moderate, or severe impairment post-stroke, providing ranges of GBH and lesion values associated with each level of recovery. Decision tree predictive models will be validated on a separate dataset from a recently completed observational study (STRONG with baseline clinical MRIs and 3-, 6-, and 12-month outcomes (N=488) and openly shared as an easy-to-use, downloadable software toolkit. Successful completion of this work will lay the foundation for personalized treatments based on individuals' unique brain health profiles, such as targeted pharmaceuticals or lifestyle interventions, with implications beyond stroke to other age-related diseases.
