Abstract Stroke is a disabling cerebrovascular disease that causes 5.5 million deaths each year globally. The disease progresses rapidly and irreversibly, leaving a narrow time window for intervention. Existing methods for patient selection for endo- vascular thrombectomy are suboptimal, based exclusively on simple linear threshold models applied to neuroimaging. Deep learning has shown great promise in recent years for many medical applications. We believe that it can be used to integrate imaging and non-imaging data in a seamless and data- driven way to improve stroke triage and clinical trials. The goal of this project is to develop deep convolutional neural network approaches to the initial MR and CT imaging, the most commonly performed stroke imaging protocol in acute ischemic stroke patients, and to combine this with non-imaging clinical information. We will train networks to predict the most likely final tissue and clinical outcomes under 2 extreme conditions (major reperfusion and minimal reperfusion) to estimate the treatment effect at the individual level. Next, we use the methods and learning from this first study to train deep learning models without using contrast perfusion imaging, which will improve safety, cost, and time-to-treatment. Finally, we will test the generalizability and explainability of these AI methods in external cohorts which differ in terms of population and scanner types, including testing on data from mobile CT scanners. Accomplishment of these aims will fundamentally shift the acute stroke paradigm beyond the relatively simplistic mismatch concept and replace it with a data-driven method that takes into account the immense amount of imaging and clinical data that can be brought to the stroke decision-making process. The methods developed will improve long-term outcomes and reduce of the cost of stroke care worldwide.