A study of spatial feature conservation in reduced channels of EEG-fNIRS based BCI using Deep Learning

The state of art Hybrid Brain Computer Interface (BCI) have shown improved classification of mental states either by combining different modalities or by choosing a combination of BCI activation tasks. Among these, the classification of motor imagery/executions of contralateral and ipsilateral data of upper arm is found challenging due to its spatial adjacency and retention of these spatial features. The proposed work uses a Hybrid BCI dataset acquired using EEG and fNIRS for upper limb movement (Right hand/ Left Hand, Right Arm/Left Arm). The electrode positioning is along the motor cortex and previous deep learning studies have shown that a good accuracy can be obtained without any channel selection. Hence the current study is to apply a combination of deep learning methods to the data which was halved into two without using channel selection algorithms. The model was evaluated for both set of channels using F1-score, Precision and Recall with an accuracy of 90%. This investigation shows that all the channels of the studied dataset contained inter-related spatial information. Also, the problem of long term EEG/fNIRS recording can be addressed using this study, if the total number of channels can be used in two halves by switching the channels after the minimum efficient time of recording.