Decoding Motion Trajectories in an Upper Limb BCI: Linear Regression vs Deep Learning

Non-invasive electroencephalogram (EEG) based brain-computer interfaces (BCI) aim to achieve control using only brain signals. Motion trajectory prediction (MTP) is a method that can be used for translating imagined three-dimensional (3D) movement to virtual limb control. This process can be enhanced in an experimental setup using 3D embodied visual feedback along with more advanced decoding approaches. Our previous studies have used multilinear regression (mLR) as a method of decoding imagined limb movement, achieving a correlation but lack the capacity to identify non-linear relationship in the data and depend on optimised features through a grid search or other approaches. Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM)-based decoders have achieved higher correlation due to their ability to address the shortcoming listed above with extensive hyperparameter optimisation. This work presents an experimental setup for an online MTP BCI using 2D and 3D visual feedback and focuses on comparing both mLR and CNN LSTM decoding methods. Preliminary results in this pilot study demonstrated that CNN LSTM significantly improves decoding performance with an average improvement of r=0.4(p < 0.01).