Searching for Causes of Poor Classification Performance in a Brain-Computer Interface Speller

The BCI Speller is a machine that flashes different letters on a screen and collects EEG data from the viewer, then analyzes the data to determine what letter the viewer was concentrating on. This allows a user to spell words without any body movements, which is useful for cases of heavy paralysis where speech is inhibited. The neurological event that this machine searches for is the P300, a pulse of electrical activity throughout the brain about 300 milliseconds after an unlikely event. However, the P300 has been noted as a varying event, and likewise the effectiveness of the Speller is highly variable: some tests will result in perfectly accurate spelling, while another test with the same user and conditions can be so inaccurate as to guess no letters correctly. The focus of this research was to search for a fundamental difference between highscoring and poor-scoring Speller tests. This could point towards a physical cause of inaccuracy, or could guide the way for a classification system that performs better in the case of a poor session. Two signal features were determined particularly likely to be the cause of classification error – variance in the latency time of the P300 and the occurrence of large-scale noise artifacts. It is difficult to mathematically define the quantity of either of these features; instead, classification methods that account for these features were designed and implemented on test sessions.