Application of Neural Networks for Decoding Kinematic Motion Parameters by Ecog Signals

In this research paper we try to improve the quality of decoding electro - corticographic (ECoG) signals recorded from the sensorimotor cortex into limb kinematics parameters using deep learning methods. We demonstrate that convolutional neural networks are able to achieve significantly higher accuracy characteristics. We managed to build an architecture based on deep convolutional networks, which provided a 20\% higher accuracy of decoding the finger trajectory in most of our experiments compared to both improved classical regression methods and neural networks architectures proposed in other scientific papers. Visualization of neural network gradients (so-called sensitivity analysis) explicitly shows that the brain electrical signals obtained from the immediate future of performed finger motion have more predictive power in comparison to the signals preceding the movement. This fact can be explained by arrang - ement of the grid electrodes and to the dominance in signals capturing the proprio - ceptive information recorded by the electrodes. In addition to decoding tasks, neural network architectures can be used for knowledge extraction. For instance, one of its applications is pattern detection of movements representation in the cerebral cortex. However, the interdependence of signals recorded by ECoG electrodes and the activity, which is unrelated to the motor function, complicate the interpretation of sensitivity analysis results. In order to solve this problem, we use the indepen - dent component analysis (ICA) for the first step of preprocessing ECoG signals before putting the dataset to the input of deep learning architectures. The results show that topographies of virtual channels (linear combinations of real ECoG leads with coefficients found by ICA) with the greatest absolute gradient values can be well-interpreted from the physiological point of view.