The effect of limiting trial count in context aware BCIs: A case study with language model assisted spelling

Abstract

Deflections in recorded electroencephalography (EEG) in response to visual, auditory or tactile stimuli have been popularly employed in non-invasive EEG based brain computer intefaces (BCIs) for intent detection. For example, in an externally stimulated typing BCI, an accurate estimate of the user intent might require long EEG data collection before the system can make a decision with a desired confidence. Long decision period can lead to slow typing and hence the user frustration. Therefore, there is a trade-off between the accuracy of inference and the typing speed. In this manuscript, using Monte-Carlo simulations, we assess the speed and accuracy of a Language Model (LM) assisted noninvasive EEG based typing BCI, RSVPKeyboardTM, as a function of the maximum number of repetitions of visual stimuli sequences and the intertrial interval (ITI) within the sequences. We show that the best typing performance with RSVPKeyboardTM can be obtained when ITI=150 ms and maximum number of allowed sequences is 8. Even though the probabilistic fusion of the language model with the EEG evidence for joint inference allows the RSVPKeyboardTM to perform auto-typing when the system is confident enough t o make decisions before collecting EEG evidence, our experimental results show that RSVPKeyboardTM does not benefit from auto-typing.