Decoding both intention and learning strategies from EEG signals

Abstract

Despite the fact that a majority of Brain-Computer Interface (BCI) studies have focused on decoding signals related to movement, decoding intention underlying movement might be equally important. Fundamental properties of electroencephalography (EEG) constrain temporal resolutions for directly decoding movement kinematics, so estimating intention pertaining to movement, which would help us predict both long and short term human behaviors, may be an alternative solution for overcoming this issue. To estimate movement intention, recent studies found it helpful to consider hierarchical control between two distinctive learning strategies in reinforcement learning: A goal-directed and a habitual strategy. In the previous study, we suggested a proof of concept, called a model-based BCI framework, that can distinguish different learning strategies from electroencephalography (EEG) data. To advance this concept, we proposed intention decoders based on simple long short-Term memory models (LSTM). To test the effect of intention estimation on prediction performance, we trained two versions of intention decoders: one with and the other without making use of underlying learning strategies decoded by the model-based BCI. The simulation results demonstrated that estimation performance of intention decoders with model-based BCI is significantly better (84%) than the one without model-based BCI (77%). We argued that by using model-based BCI, we can not only estimate learning strategy but also improve decoding performance of movement intention with high accuracy.