Improving Motor Imagery-Based Brain-Computer Interfaces with Simple EEG Data Augmentation Algorithms: A Comparative Analysis

One of the primary challenges in motor imagery (MI) based brain-computer interfaces (BCIs) is the extensive calibration time required for individual users. This paper systematically reviews and compares various straightforward EEG data augmentation techniques aimed at reducing calibration time and enhancing model performance. We explored several algorithms, including segmentation and recombination in the time-frequency domain using short-time Fourier transform (TFSR), time warping (TW), noise injection (NI), and naïve transfer learning (NTL). Additionally, we introduced a novel algorithm called time instance segment concatenation (TISC), which merges EEG segments from periods immediately before and after the traditional motor imagery segments to capture a broader range of neural dynamics. Our experimental results on the BCI Competition IV dataset 2a indicated that NTL was the top-performing algorithm, achieving average accuracy improvements of 4.9%, 3%, 1.7%, and 3.2% over the baseline model without any data augmentation when 5, 10, 15, and 20 training trials per class were available, respectively. The proposed TISC algorithm showed potential, particularly in scenarios with moderate training trials, e.g., achieving an average accuracy increase of 1.5% over the baseline when 10 training trials per class were available. These findings suggest that TISC could be a useful addition to EEG data augmentation methods to improve the efficiency and usability of MI-based BCIs.