Towards Applicability of Motor Imagery BCI: Study on Artificial EEG Data Generation Methods for Calibration Time Reduction

• Published in: 2020 3rd International Conference on Biomedical Engineering (IBIOMED) pp. 1 - 6
• Main Authors: Anindya, Sinantya Feranti, Darmakusuma, Reza
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06-10-2020
• Subjects: artificial data generation; brain-computer interface; calibration time reduction; Electroencephalography; Empirical mode decomposition; Fitting; Image segmentation; motor imagery; Time measurement; Time-frequency analysis; Training
• DOI: 10.1109/IBIOMED50285.2020.9487602

This work explores applicability of artificial EEG data generation (ADG) methods to reduce calibration time of motor imagery-based BCI, which will be used for reducing the calibration time of a real-time MI BCI in future work. Three methods are explored: data generation based on time segments, time-frequency segments, and empirical mode decomposition. To compare the methods, the left-hand and right-hand data of nine subjects from the BCI Competition IV 2a dataset are used as experiment material. First, each training trial is preprocessed and divided into smaller components using the selected methods. Second, an artificial trial for one class is created by combining several randomly chosen components from the targeted class. The process is repeated 100 times for each class, resulting in 200 artificial trials in total. Afterwards, a new training dataset combining the original and artificial trials is used for fitting the classifier, which consists of CSP and LDA. The performance of each data generation method is measured based on two parameters: time elapsed for ADG process and classifier accuracy. Overall, ADG based on time segments performed the fastest in terms of decomposition time (0.0255 s) and data construction time (0.1969 s) taken. However, while artificial trials from time segments and time-frequency segments are able to improve classifier accuracy when small training sample is used (73.73% and 75.16% versus 68.45% for 24 original training trials), baseline classifier without artificial trials still performs better with more original training trials.