Computational neural network provides naturalistic solution for recovery of finger dexterity after stroke

Abstract

Finger dexterity is a fundamental movement skill of humans and the ability to individuate fingers imparts high motor flexibility. Disruption of dexterity due to brain injury reduces quality of life. Thus, understanding the neurological mechanisms responsible for recovery is critical to effective neurorehabilitation. Two neuronal pathways have been proposed to play crucial roles in finger individuation: the corticospinal tract, originating from primary motor cortex and premotor areas, and the subcortical reticulospinal tract, originating from the reticular formation in the brainstem. Finger individuation in patients with lesions to these pathways may recover. However, it remains an open question how the cortical-reticular network reorganizes and contributes to this recovery following a stroke. We hypothesized that interactive connections between cortical and subcortical neurons reflect dynamics appropriate for generating outgoing commands for finger movement. To test this hypothesis, we developed an Artificial Neural Network (ANN) representing a premotor planning input layer, a cortical layer including excitatory and inhibitory neurons and, a reticular layer that control motoneurons eliciting unilateral flexion of two fingers. The ANN was trained to reproduce "normal" activity of finger individuation and strength. Analysis of the trained ANN revealed that the natural dynamical solution was a near-linear relationship between the force of the instructed and uninstructed finger, resembling individuation patterns in humans. A simulated stroke lesion was then applied to the ANN and the resulting finger dexterity was assessed at multiple stages post stroke. Analysis revealed: (1) increased unintended force produced by uninstructed fingers (i.e., enslaving) and (2) weakening of the force in the instructed finger immediately after stroke, (3) improved finger control during recovery that typically occurs early after stroke, and (4) association of this behavior with increased neural plasticity of the residual neurons, as reflected by strengthening of connectivity weights between premotor and focal cortical excitatory and inhibitory neurons, but reduction in connectivity in shared cortical neurons. Interestingly, the network solution predicted that the reticulospinal pathway also contributed to the improved behavior. Lastly, the ANN also predicts the effect of cortical lesion size on finger individuation. Our model provides a framework by which to understand a number of experimental findings. The model solution suggests that a key mechanism of finger individuation is establishment of an interactive relationship between cortical and subcortical regions, appropriate to produce desired finger movement.