Classification of Raw Spinal Cord Injury EEG Data Based on the Temporal-Spatial Inception Deep Convolutional Neural Network

Today, decoding electroencephalography (EEG) data is an important and efficient achievement in neuroscience and Brain-Computer Interface (BCI) systems. So that improving the performance of BCI systems can be effective in helping spinal cord injury (SCI) patients to perform daily tasks and rehabilitation work. In contrast, some of the challenges that make this path uneven include low signal-to-noise ratios, various artifacts in EEG signals, Low-performance methods based on handcraft feature extraction, the need for expert people, and differences between different people. However, today the development of deep learning methods and the availability of EEG datasets have reduced the existing limitations. Deep learning frameworks provide an end-to-end approach with the help of various structures that, by extracting automatic features from raw data, they try to reduce the limitation of decoding EEG data and increase the performance of BCI system as much as possible. One of the most practical and effective structures in this field is deep convolutional models. These models learn different features at different scales by taking raw EEG data in various forms at their input without any restrictions. In this paper, we present two deep convolutional models, TSTI-Net and TSI-Net, with the help of designed Inception modules. Taking advantage of the Large and Tiny Inception modules and extracting spatial-temporal features, these two models achieve good results in classifying raw EEG data of spinal cord injury patients. Each of the two models consists of three parts. The first and second parts have the same structure in both models. In these parts, temporal and spatial features are extracted with the help of 2D-temporal and Depthwise convolutions. In the third part, two different Inception modules for each model are used. The two models TSTI-Net and TSI-Net, classify the EEG data of spinal cord injured persons into six classes with 69.29% and 71.45% classification accuracy, respectively. Also, the results expressed for the two proposed models compared to the three famous models EEGNET, ShallowConvNet, and DeepConvNet in the classification of raw EEG data show the good performance of the designed structures TSI-Net and TSTI-Net.