Exploring Two Novel Features for EEG-based Brain-Computer Interfaces: Multifractal Cumulants and Predictive Complexity

TL;DR

This paper introduces two novel EEG features, multifractal cumulants and predictive complexity, which improve BCI classification performance when combined with traditional band-power features.

Contribution

The paper presents two new features for EEG-based BCI, demonstrating their effectiveness and synergy with existing features in motor-imagery tasks.

Findings

Novel features improve classification accuracy

Combining features yields best performance

Features outperform traditional band-power in some cases

Abstract

In this paper, we introduce two new features for the design of electroencephalography (EEG) based Brain-Computer Interfaces (BCI): one feature based on multifractal cumulants, and one feature based on the predictive complexity of the EEG time series. The multifractal cumulants feature measures the signal regularity, while the predictive complexity measures the difficulty to predict the future of the signal based on its past, hence a degree of how complex it is. We have conducted an evaluation of the performance of these two novel features on EEG data corresponding to motor-imagery. We also compared them to the most successful features used in the BCI field, namely the Band-Power features. We evaluated these three kinds of features and their combinations on EEG signals from 13 subjects. Results obtained show that our novel features can lead to BCI designs with improved classification performance, notably when using and combining the three kinds of feature (band-power, multifractal cumulants, predictive complexity) together.