Project Summary The large variability in lesions, impairment, and responsiveness to therapy following stroke has hindered the development of principled and effective approaches to neurorehabilitation of the upper extremity (UE). Here, building on our multidisciplinary expertise in large-scale neurorehabilitation studies, brain imaging, and data science, as well as our established collaborations, we propose a novel self-improving personalized rehabilitation system based on a continuous loop between data, algorithm, and treatment. In Aim 1, we will create and simulate a self-improving algorithm for personalized UE rehabilitation, based on the state-of-the-art and theoretically-sound “Bayesian contextual bandits” algorithm. A predictive dynamics model will generate time- varying predictions of the UE functional outcome (the ARAT) at 3 months post-stroke in response to training and based on the patient baseline clinical and neural variables that modulate the effects of treatment, and on prior knowledge from similar patients. Second, a sequential-decision bandit algorithm will use the model’s predictions for different (simulated) doses to select the weekly dose that is predicted to maximize the ARAT at 3 months post-stroke, given constraints on the feasible doses. In Aim 2, we will conduct a cohort study of 400 individuals with sub-acute stroke. The detailed treatment, neural, and clinical data generated by this study will allow us to identify the clinical and neural modulators of sensorimotor training as well as the feasible doses in sub-acute stroke. In particular, treatment data will include the schedule of arm movements recorded via an exoskeleton during mechanized adjuvant therapy and the number of hand movements both during and outside conventional therapy recorded with a wrist-worn sensor. In Aim 3, we will perform a proof-of-concept study of the self-improving precision neurorehabilitation system, with 150 participants in the sub-acute phase post-stroke. The algorithm of Aim 1 will determine the optimal weekly doses based on previous measurements of the actual doses, the measured ARAT, and the participant clinical and neural context. This Aim will start in year 3 after the data from 200 participants in Aim 2 have been collected to update the algorithm of Aim 1. The data generated by this Aim will continuously be used to update the model of Aim 1, yielding better predictions and thus more effective schedules of weekly doses. We will therefore test the hypothesis that the change in ARAT from baseline to 3-month will be significantly greater for each new group of 50 participants compared to the previous group. Upon completion of this research, we will have captured a comprehensive picture of rehabilitation treatment parameters that most improve UE function given the heterogeneous profiles after stroke. The resulting unique database of 550 participants, which will be made publicly available, is expected to become a reference in neurorehabilitation...