Simulation of an individual with motor disabilities by a deep reinforcement learning model

Abstract

We have developed a new neural network model that simulates how the central nervous system (CNS) governs neural motor sensors. Our model uses reinforcement learning and transfer entropy to compare healthy individuals’ learning with those who have motor impairments. Our aim is to study effective connectivity and identify differences in information transmission between the two groups. By analyzing effective connectivity, we have identified patterns arising from zero and non-zero transfer entropy values. These patterns are mainly due to factors such as the level of connectivity, the neural network’s learning time, and the rules guiding the model training process. However, when these patterns appear in the input and output layers, they are not necessarily critical in distinguishing between a healthy network and an impaired one since these layers interact directly with the environment. On the other hand, our model can replicate neurobiological factors linked to neuronal injury, indicated by these patterns, when they appear in the intermediate layers. We found that healthy networks had more neurons involved in sensorimotor communication compared to diseased networks. This may be due to an increase in disease-related glial cells, which led to reduced effective connectivity between CNS regions. Additionally, we observed some patterns for the loss or gain of entropy; these patterns show the sensitivity of our model to the number of delays we take into account for the effective connectivity to ensure that the information transfer within the model behaves as a natural process.