Leveraging Unsupervised Learning Methods to Improve Imagined Speech EEG-BCIs

A BCI is a direct communication pathway between the brain and an external device with the goal of providing communication for individuals with motor impairments. However, a considerable percentage of subjects cannot modulate their brain rhythms to communicate using BCI. From a machine learning standpoint, the instability of brain signals poses a critical challenge, which is the validity of the training data labels. Furthermore, the data generated in one’s mind cannot be directly evaluated. One potential solution is the inclusion of unsupervised methods in the learning processes of a BCI. This research proposes two methods based on unsupervised learning to relabel suspicious data, and to capture within class variations of data. These methods were tested in the development of imagined speech EEG-BCIs for 12 participants and significantly improved the classification performance for 7 of them. Future works are encouraged to use these methods to individualize the design of BCI protocols.M.A.S