Methods for Performance Improvement in Recurrent Rehabilitative Brain Computer Interface Applications^*

Xin Zhang,Carlo Menon

DOI: https://doi.org/10.1109/CBS55922.2023.10115382

2023-01-01

Abstract:EEG based brain computer interfaces (BCIs) have been proven to be effective in priming recovery in stroke rehabilitation. Due to the non-stationary nature of the EEG signal, cross-sessional BCI applications generally suffer from performance issues, which not only reduce the training efficiency but also deteriorate the entire system stability, especially with increased false positive outputs. However, it is very inefficient to train BCI models everyday for long-term recurrent applications like rehabilitation training. In this paper, we preliminarily investigated the feasibility of two methods for suppressing false positive outputs and fast re-calibrating online application in long-term recurrent scenarios. Biased-classification and moving-average methods were proposed to mitigate the gap between offline model generation and online application. The proposed two methods were evaluated on EEG data from one participant with chronic stroke during six weeks of rehabilitation training using BCI-controlled exoskeleton. Two proposed configuration methods were introduced and evaluated on the recorded EEG data in a simulated online paradigm. The result showed that the proposed online classification filtering methods were able to improve classification performance by considerably suppressing false positive outputs in rehabilitation applications. The proposed methods contributed in improving the online BCI performance and streamlining the re-calibration BCI process in recurrent long-term BCI applications.