Prediction error decoding from EEG signals in a realistic environment

Abstract

Recent studies have revealed that decoding underlying learning strategies helps to decode the corresponding behavior or movement. This is because high-level cognitive states such as intention, learning strategies, or prediction errors (PEs) always precede and underlie every form of movement. However, Brain-Computer Interface (BCI), which serves as a communication channel between the brain and the machine, has focused on motor function-related EEG signals. Because motor function-related EEG signal delivers information only about which and how muscles are moved, it may not suitable for decoding high-level cognitive functions or complex movements. Related works have relied on a limited experimental environment, such as a 2-stage Markov Decision Task (MDT), making it hard to achieve realistic PE signals. To settle this issue, this study aims to decode the high-level cognitive state PE signal in a realistic environment. Specifically, we propose a novel realistic task paradigm with a high variability of PEs and implement the EEG-based PE decoder. As a proof of concept, we tested our model on the 2-stage MDT dataset obtained in the previous study and achieved an accuracy of 61.3 % for SPE and 58.4 % for RPE. Finally, we achieved an accuracy of 97.4 % for RPE in the realistic environment we suggested here.