Spatiotemporal Analysis of Time Window Length in Multi-Directional Motor Imagery Classification Using Emotiv Insight

This paper proposes to examine the feasibility of classifying multi-directional motor imagery tasks by using the low-channel electroencephalography signal with a consumer-grade headset, thus trying to fulfill the demand for more practical and affordable brain-computer interfaces. EEG signals were recorded from 53 participants under a video-guided MI paradigm, with preprocessing including channel selection, average referencing, band-pass filtering, ICA-based signal inspection, and temporal windowing to construct sequential inputs for deep learning models, while participants maintained central gaze fixation throughout all tasks to minimize potential directional eye-movement confounds. Signal quality was assessed using independent component analysis prior to epoch extraction. The proposed method is tested with a convolutional neural network, a long short-term memory network, and a combined CNN and LSTM network for classifying the four motor imagery direction tasks. The CNN–LSTM achieved the highest mean classification accuracy (87.14%) across four directional classes (left, right, front, back), outperforming the CNN (83.79%) and LSTM (79.74%), with the best Information Transfer Rate of 57.5 bits/min. Results confirm that joint spatial–temporal modeling significantly enhances MI decoding under low-channel constraints, demonstrating the viability of consumer-grade EEG for scalable, real-world BCI applications.