Project Summary Poststroke cognitive impairment (PSCI) was found to be common in various research studies. PSCI is ideally recognized through cognitive screening and test but they are not standard clinical practices and hence stroke recovery and prevention of recurrent strokes may be undermined by concurrent but poorly recognized cognitive issues, e.g., patient compliance to follow blood pressure control medication may be poorer among those with PSCI. Therefore, a significant unmet need for optimizing poststroke care is to recognize patients at high risk of PSCI to tailor for them an appropriate stroke recovery and recurrent stroke prevention strategy. With many of the plausible determinants of PSCI being available in electronic health record (EHR) systems, machine learning (ML) methods to process routine clinical data to predict risk of PSCI is highly feasible. We propose to combine a large retrospective dataset from EHR and a smaller prospective dataset with more accurate ascertainment of PSCI based on purposefully administered cognitive tests, serving as gold-standard. The necessity of prospective cognitive tests to accurately ascertain PSCI further allows us to explore biological and physiological variables related to pathologies of Alzheimer disease and related dementia (ADRD). We will pursue three specific aims: 1) Learn to predict PSCI using routine neuro images and EHR data from large clinical cohorts; 2) Use prospective data to adapt and validate models learned from existing clinical cohorts; 3) Phenotype PSCI with cognitive tests, physiological, and biological metrics one-year poststroke. Prediction of PSCI could aid optimizing stroke recovery and recurrent stroke prevention strategies. Our proposed novel physiological and biological metrics have the potential to further improve PSCI prediction and characterize PSCI granularly with the consideration of cerebrovascular and neurodegenerative underpinnings.