Exploring the Most Significant Features for ErrP Detection Through Statistical Analysis

Abstract

Recently, electroencephalographic (EEG) signals have been used to enhance human-robot interaction (HRI), particularly through error-related potentials (ErrPs), a type of Event Related Potential (ERP). ErrPs provide feedback on mismatches between user expectations and robot behavior. Correct classification of ErrPs is crucial and depends on reliable extraction and selection of signals features. This work evaluates a comprehensive list of features through statistical analysis to identify the most relevant ones for improving classification.

Introduction

An ERP is a brain response to a specific stimulus, such as visual or auditory cues, and is not linked to correctness. Instead, the ErrPs are involuntarily evoked when a person perceives an unexpected error or unexpected actions during an interaction task. ErrPs are crucial for error detection as they cannot be masked and are unintentionally evoked, making them reliable signals for error detection during interactions [1]. They are characterized by a specific waveform lasting 600-800 ms, though experimental setup can affect latency. ErrPs have been used in brain-robot interface (BRI) and brain-computer interface (BCI), for both online and offline control, following a similar process: acquiring EEG signals, classifying them through feature extraction and selection, and providing feedback control to the robot.

Methodology

This study evaluates an extensive list of potential features through statistical analysis to gauge their discriminative power for ErrP analysis, aiming to reduce the number of features used for classification. The dataset [2] comprises recordings from 11 participants engaged in two different interaction scenarios: cursor and robot. In both scenarios, participants respond to visual stimuli by pressing corresponding keys, prompting movement of a cursor or humanoid robot. An ErrP is evoked if the movement does not correspond to the user’s input, with error probabilities of 20% and 50% for the cursor and robot scenarios, respectively. After signal processing, relevant features extracted included temporal, frequency, statistical, and Wavelet coefficients. The statistical analysis employed two non-parametric tests (Kolmogorov-Smirnov and Wilcoxon rank sum) and one parametric test (t-test), preceded by checks for normal distribution using the Shapiro-Wilk and Lilliefors methods.

Results

Overall, the outcome of our study shows that some parameters have relevant importance compared to others. The features relevant after the statistical analysis considering both scenarios are prominence, time and amplitude of the maximum and minimum peak, the RMS, the standard deviation, the maximum value and the frequency of FFT, the median frequency, the power band-width, the SNR, the Wavelet coefficient 8,9,10,11,10 of the [4-8] Hz band and the coefficient 6 of the [2-4] Hz band. We think that using these features the binary classification could be improved: future work will aim at validating this hypothesis.

References

[1] X. Wang, H.-T. Chen, Y.-K. Wang, and C.-T. Lin, “Implicit robot control using error-related potential-based brain–computer interface,” IEEE Transactions on Cognitive and Developmental Systems, vol. 15, no. 1, pp. 198–209, 2023.

[2] S. K. Ehrlich and G. Cheng, “A feasibility study for validating robot actions using eeg-based error-related potentials,” International Journal of Social Robotics, vol. 11, 04 2019.