Modeling the Relationship Between Cognitive State and Task Performance in Passive BCIs using Cross-Dataset Learning

New research and development efforts are highlighting the ways in which electroencephalogram-based (EEG) brain-computer interface (BCI) technology can be used to improve the quality of life for healthy individuals. One such application incorporates cognitive state monitoring into passive BCI (pBCI) systems. Among the challenges facing this development, a significant barrier to adoption is the time-intensive calibration typically needed to tune the system to account for variations in neural activity patterns. An open research question is understanding the relationship between underlying user state and user performance in real-world situations and environments. However, user states are often derived and defined as a function of observed user performance for a particular analysis. Understanding the relationship between user state and user performance ideally requires the definition of user state be independent of observed user performance. This work represents our initial steps towards this goal by using cross-dataset learning, where we define user state from a dataset recorded in a highly-controlled experiment, building a subject-independent pBCI model to predict user state using deep learning approaches, and applying this pBCI model to analyze user performance in a new, unobserved dataset. We show that user performance varies smoothly across a continuum of pBCI model outputs. Our results highlight a promising approach for dealing with one of the major hurdles in the development of BCI systems for healthy users.