Identification of cerebral cortices processing acceleration, velocity, and position during directional reaching movement with deep neural network and explainable AI

Cerebral cortical representation of motor kinematics is crucial for understanding human motor behavior, potentially extending to efficient control of the brain-computer interface. Numerous single-neuron studies have found the existence of a relationship between neuronal activity and motor kinematics such as acceleration, velocity, and position. Despite differences between kinematic characteristics, it is hard to distinguish neural representations of these kinematic characteristics with macroscopic functional images such as electroencephalography (EEG) and magnetoencephalography (MEG). The reason might be because cortical signals are not sensitive enough to segregate kinematic characteristics due to their limited spatial and temporal resolution. Considering different roles of each cortical area in producing movement, there might be a specific cortical representation depending on characteristics of acceleration, velocity, and position. Recently, neural network modeling has been actively pursued in the field of decoding. We hypothesized that neural features of each kinematic parameter could be identified with a high-performing model for decoding with an explainable AI method. Time-series deep neural network (DNN) models were used to measure the relationship between cortical activity and motor kinematics during reaching movement. With DNN models, kinematic parameters of reaching movement in a 3D space were decoded based on cortical source activity obtained from MEG data. An explainable artificial intelligence (AI) method was then adopted to extract the map of cortical areas, which strongly contributed to decoding each kinematics from DNN models. We found that there existed differed as well as shared cortical areas for decoding each kinematic attribute. Shared areas included bilateral supramarginal gyri and superior parietal lobules known to be related to the goal of movement and sensory integration. On the other hand, dominant areas for each kinematic parameter (the contralateral motor cortex for acceleration, the contralateral parieto-frontal network for velocity, and bilateral visuomotor areas for position) were mutually exclusive. Regarding the visuomotor reaching movement, the motor cortex was found to control the muscle force, the parieto-frontal network encoded reaching movement from sensory information, and visuomotor areas computed limb and gaze coordination in the action space. To the best of our knowledge, this is the first study to discriminate kinematic cortical areas using DNN models and explainable AI