Level-wise Subject adaptation to improve classification of motor and mental EEG tasks

Classification of various cognitive and motor tasks using electroencephalogram (EEG) signals is necessary for building Brain Computer Interfaces (BCI) that are noninvasive. However, achieving high classification accuracy in a multi-subject multitask scenario is a challenge. A noticeable reduction in accuracy is observed when the subjects between train and test are mismatched. Drawing a similarity from speaker adaptation approaches in speech, we propose a method to perform subject-wise adaptation of EEG in order to improve the task classification performance. A Common Spatial Pattern (CSP) approach is employed for feature extraction. Gaussian Mixture Model (GMM) based subject-specific models are built for each of the tasks. Maximum a-posterior (MAP) adaptation is performed, and an absolute improvement of 1.22-7.26% is observed in the average accuracy.