Fig. 4

Examples of computational neurorehabilitation approaches and results. a A key output of the Han et al. model [112] is the predicted spontaneous use of the impaired hand, shown here as a percent of all movement trials in a bimanual choice task. Each curve represents the evolution of spontaneous use given the number of rehabilitation practice trials, shown as a number on the far right of each curve. Spontaneous use increases only when enough rehabilitation practice trials are performed to reach a threshold. From [112]; used with permission. b A key output of the Casadio et al. model [56], which used data from a robotic therapy trial, is that the retention parameter in the model, measured through a trial-to-trial analysis, predicts the change in Fugl-Meyer score at 3 months for these chronic stroke participants. c The Reinkensmeyer et al. model [136] assumes that wrist force is produced by the summed effect of corticospinal cells targeting motor neuronal pools. Each cell contributes an incremental force proportional to its firing rate, up to a saturation level. Cell firing rate changes by a random amount from trial to trial; activation patterns that produce more force are remembered for future use, thus implementing a reinforcement learning paradigm. d In the Reinkensmeyer et al. model, the probability that a single neuron will contribute to an increase in force on a new trial depends on whether the neuron is strongly or weakly connected to the motor neuronal pool. Strongly connected cells have a greater probability of producing a larger increase. In addition, when cells become saturated, they can only decrease force production. Thus, an increasing number of saturated cells increasingly blocks further optimization, leaving a residual capacity for further increases in force