The Self-Paced Graz Brain-Computer Interface: Methods and Applications

Reinhold Scherer, Alois Schlögl, Felix Lee, Horst Bischof, Janez Jansa, Gert PfurtschellerComput. Intell. Neurosci 2007 · DOI · OpenReview

TL;DR

This paper introduces the "self-paced" Graz Brain-Computer Interface (BCI), enabling users to control applications by intentionally thinking about motor imagery without external cues. The system effectively handles common artifacts like eye movements (EOG) and muscle activity (EMG), paving the way for more natural and intuitive BCI applications.

What Problem Does This Paper Address?

Brain-Computer Interfaces (BCIs) offer a way for individuals to control external devices using their brain activity, bypassing traditional motor pathways. However, existing BCIs often rely on external cues (like a flashing letter on a screen) to signal when the user should generate a specific brain pattern. This "cue-based" approach limits the naturalness and speed of interaction. The authors address the challenge of creating a "self-paced" BCI, where the user can initiate control commands at any time, mimicking natural human interaction. This requires the BCI to not only detect intended brain signals but also to distinguish them from ongoing, non-control brain activity, all while dealing with common signal artifacts.

Methods

The Graz BCI employs several key methods:

  • Self-Paced Operation: The core innovation is the ability to differentiate between "intentional control" (IC) signals generated by motor imagery and "non-control" (NC) states. This is achieved by training two classifiers: one (CFRIC) to detect motor imagery-related brain activity and another (CFRMI) to determine the specific type of motor imagery. The system only acts when CFRIC detects activity, and CFRMI then identifies the command.
  • EMG Artifact Detection: To prevent muscle activity from triggering unintended commands, an inverse filtering method is used. This method estimates an autoregressive (AR) model of the EEG. When high-frequency artifacts (like EMG) are present, the prediction error of this model increases significantly, allowing for their detection and reporting to the user.
  • EOG Artifact Reduction: Eye movements (EOG) are a common source of noise in EEG. The paper utilizes a method that models EOG as a linear superposition of real EEG signals and three spatial EOG components. By estimating the weighting coefficients, EOG artifacts are effectively subtracted from the EEG signal, reducing their impact by approximately 80%.
  • Feature Extraction: Band power (BP) features from specific frequency bands (10-12 Hz and 20-25 Hz) of the EEG are used as input for the classifiers. These features are extracted from motor imagery trials and also from artifact-free EEG segments for the non-control state.
  • Classification: Fisher's linear discriminant analysis (LDA) is used for classification, combined with a majority vote for CFRMI. For CFRIC, a distinction-sensitive learning vector quantization (DSLVQ) method is employed to identify the most discriminative features.

Datasets

The paper describes experiments conducted with three healthy subjects. The datasets used were collected during these experiments, specifically:

  • EEG and EOG data: Collected for setting up EMG detection and EOG reduction. This involved specific protocols of eye movements (blinking, rolling, horizontal/vertical movements) and resting periods.
  • Motor Imagery (MI) data: Collected during cue-based feedback training sessions. This data was used to train the CFRMI classifier.
  • Self-paced BCI data: Collected during the operation of the "freeSpace" virtual environment and the "Brainloop" interface. This data was used to evaluate the performance of the self-paced BCI system.

No pre-existing public datasets are explicitly mentioned as being used.

Key Insights and Ideas

  • Self-Paced Control is Achievable: The most significant insight is the successful implementation of a BCI that allows users to initiate control commands naturally, without external cues, by distinguishing intentional brain activity from background brain states.
  • Artifact Handling is Crucial for Practicality: The paper highlights the importance of robust methods for detecting and reducing common artifacts like EMG and EOG. Without these, the reliability and usability of BCIs would be severely compromised.
  • Simplicity Can Be Effective: The system uses only three bipolar EEG channels and relatively straightforward artifact reduction and feature extraction techniques, demonstrating that complex setups are not always necessary for effective BCI operation.
  • User Adaptation is Key: While the BCI system is designed to adapt to the user, the paper also notes that users learned to effectively generate the required mental states for control, suggesting a bidirectional learning process.

Key Results

  • EMG Artifact Detection: The EMG detection method was highly effective, classifying less than 0.9% of EEG samples as EMG artifacts during training.
  • EOG Artifact Reduction: The EOG reduction method successfully reduced approximately 80% of EOG artifacts.
  • Classification Accuracy: The LDA classifiers trained for self-paced control achieved classification accuracies of 77%, 84%, and 78% for the three subjects.
  • Application Performance:
    • freeSpace VE: All three subjects successfully navigated the virtual environment and collected coins. Two out of three subjects managed to collect all three coins within the time limit.
  • Brainloop Interface: One subject successfully operated Google Earth using the Brainloop interface, demonstrating the system's applicability to complex software.

Limitations and Future Work

  • Long-Term Stability: The authors acknowledge that the long-term stability of the EMG and EOG artifact handling methods is an open issue.
  • Threshold and Transition Time Optimization: The optimal detection threshold (THIC) and transition time (tr) were determined empirically based on subject feedback, which is noted as a suboptimal solution. Future work could involve more systematic methods for parameter optimization.
  • Performance Measures for Self-Paced BCIs: The paper points out the lack of standardized benchmark tests and performance measures for asynchronous or self-paced BCIs, hindering direct comparison with other systems.
  • User Training: While subjects learned to use the system, the duration and effectiveness of training could be further investigated.

Future work could focus on developing more robust and adaptive artifact handling techniques, exploring more sophisticated feature extraction and classification methods, and establishing standardized evaluation protocols for self-paced BCIs.

Summary updated

Abstract

We present the self-paced 3-class Graz brain-computer interface (BCI) which is based on the detection of sensorimotor electroencephalogram (EEG) rhythms induced by motor imagery. Self-paced operation means that the BCI is able to determine whether the ongoing brain activity is intended as control signal (intentional control) or not (non-control state). The presented system is able to automatically reduce electrooculogram (EOG) artifacts, to detect electromyographic (EMG) activity, and uses only three bipolar EEG channels. Two applications are presented: the freeSpace virtual environment (VE) and the Brainloop interface. The freeSpace is a computer-game-like application where subjects have to navigate through the environment and collect coins by autonomously selecting navigation commands. Three subjects participated in these feedback experiments and each learned to navigate through the VE and collect coins. Two out of the three succeeded in collecting all three coins. The Brainloop interface provides an interface between the Graz-BCI and Google Earth.
BibTeX
@article{scherer2007self,
  author = {Reinhold Scherer and Alois Schlögl and Felix Lee and Horst Bischof and Janez Jansa and Gert Pfurtscheller},
  title = {{The Self-Paced Graz Brain-Computer Interface: Methods and Applications}},
  journal = {Comput. Intell. Neurosci.},
  year = {2007},
  doi = {10.1155/2007/79826},
}
Figure