Fig. 1

a General framework of computational neurorehabilitation models. Such models predict patient functional outcomes by driving computational representations of plasticity and learning with sensorimotor activity achieved in rehabilitation therapy and/or throughout the course of daily life. b Computational neurorehabilitation models presume that rehabilitation modulates both spontaneous biological recovery and motor learning, leading to improvements in both impaired limb motor control and compensatory movement strategies. Shown here is an estimate of the dose-response effect arising from additional therapy time, obtained by plotting effect sizes of 30 studies of upper and lower extremity rehabilitation therapy after stroke involving 1750 total participants as a function of the number of additional training hours ΔΤime. Note in this study there was no significant effect of the time the therapy was delivered after stroke (i.e. soon after stroke or in the chronic state). From [9]. Used with permission. c Computational neurorehabilitation models are becoming increasingly feasible in part because of a large influx of detailed kinematic data characterizing the content and outcomes of therapy, which is being obtained from robotic devices, such as Pneu-WREX shown here [218] and wearable sensors. Both individuals consented to the publication of this image. d Example of a computational neurorehabilitation model [112]. This model simplified neurorehabilitation dynamics by assuming that a reward-based learning mechanism determines the probabilities of using the impaired or unimpaired arms after stroke, and that a separate, error-based learning mechanism accounts for improvements in motor control through practice. The model predicts that if a patient reaches a threshold of recovery, then he or she will enter a positive cycle of using and further retraining their impaired arm through spontaneous activity in daily life, a prediction supported by data from the EXCITE clinical trial. Used with permission