Cross-Subject & Cross-Dataset Subject Transfer in Motor Imagery BCI systems

Abstract

Motor Imagery (MI) based Brain Computer Interfaces (BCIs) are seen as effective mechanisms for motor rehabilitation. Aside from promises of MI based BCI systems, their utility are mainly limited to laboratory-based Single-Trial EEG studies where each participant/patient undergo a long and tedious EEG data recording session to train a classifier that can accurately stratify participant's MI patterns. Session and subject transfer frameworks are considered as liable solutions for this problem. This study assess the utility of deep learning models, trained on MI patterns from other subjects enrolled in a) the same study (Cross-Subject Within-Dataset (CSWD)) or b) across multiple studies (Cross-Subject Cross-Dataset (CSCD)), for stratifying EEG patterns representing 3 different MI tasks. The aim of the study is to a) highlight the effectiveness of subject transfer in MI-based BCI studies and b) evaluate the hypothesis that “there exist a set of unique MI patterns that are known to be impacted by differences across experiment design, EEG recording equipment, phenotype of participants and many other factors, but can be used to generate universal MI-based BCI systems”. Three well known dataset of BCI Competition IV Dataset I & IIa and BCI Competition III Dataset IVa are used to assess these hypothesises. First, results from Single Trial EEG analysis is considered as base-line and later a set of experiments are conducted to assess within dataset and across dataset subject transfer (CSWD & CSCD). The results indicate the proposed subject transfer methods achieved similar or higher performances compared to state-of-the-art.