Towards a Fully Interpretable EEG-based BCI System

Fabien Lotte, Anatole Lécuyer, Cuntai Guan2010

TL;DR

This paper introduces a novel algorithm for creating Brain-Computer Interfaces (BCIs) that are not only accurate but also interpretable. By combining inverse solutions for EEG signal localization, fuzzy inference systems for rule extraction, and linguistic approximation for human-readable explanations, the BCI can explain why it makes certain decisions, opening doors for verification, improvement, and deeper understanding of brain activity.

What Problem Does This Paper Address?

Most current Brain-Computer Interfaces (BCIs), especially those relying on machine learning, operate as "black boxes." This means that while they can translate brain signals into commands, it's difficult or impossible to understand how they learned to do so. This lack of interpretability hinders several crucial aspects:

  • Verification and Improvement: It's hard to check if the BCI has learned correctly or to identify potential flaws in its design or the underlying assumptions.
  • Collaboration: Explaining the BCI's learned knowledge to domain experts (like neurophysiologists) or even non-technical users is challenging.
  • Scientific Discovery: Understanding the BCI's internal workings could provide valuable insights into brain processes related to specific mental states.

The paper aims to bridge this gap by developing an EEG-based BCI that can explicitly explain its learned decision-making process in a way that is understandable to humans.

Methods

The proposed method is a three-step pipeline:

  1. Feature Extraction based on Inverse Solutions (FuRIA): Instead of relying on surface EEG signals directly, this step uses inverse solutions to estimate brain activity in specific regions and frequency bands throughout the entire brain volume. FuRIA (Fuzzy Rule-based Inverse solution for Artifacts and Activity) specifically identifies relevant brain regions and frequency bands whose power is predictive of mental states. This provides physiologically meaningful features.
  2. Classification using Fuzzy Inference Systems (FIS): The extracted features are then fed into a Fuzzy Inference System. FIS are known for their ability to automatically learn "if-then" rules from data. These rules describe how combinations of feature values (e.g., power in a specific brain region and frequency band) map to different mental states. The output is a set of fuzzy rules.
  3. Interpretability Improvement via Linguistic Approximation: While fuzzy rules are more interpretable than standard machine learning models, they still involve fuzzy sets (mathematical representations of vagueness). To make these rules truly accessible, this step translates the fuzzy sets into simple, human-understandable words (linguistic terms like "Low," "Medium," "High"). This is achieved by defining a vocabulary of fuzzy sets and then mapping the learned fuzzy sets to the closest word in the vocabulary.

Datasets

The paper evaluates its approach using data from the BCI competition 2003, data set IV. This dataset contains EEG signals recorded from a subject performing self-paced left and right hand finger tapping. The task is to predict which hand the subject intends to use. The dataset was split into training (314 trials) and testing (100 trials) sets.

The authors also mention evaluating the approach on EEG signals from a visual spatial attention experiment, in collaboration with external researchers, but details are not reported due to space limitations.

Key Insights and Ideas

  • Bridging the Gap Between Black Boxes and Understanding: The core insight is that interpretability doesn't have to come at the cost of performance. By carefully selecting feature extraction and classification methods, it's possible to build BCIs that are both accurate and explainable.
  • Physiologically Meaningful Features: Using inverse solutions to estimate brain activity provides features that are directly linked to brain regions and frequencies, making the learned rules more grounded in neuroscience.
  • "Computing with Words" for BCIs: The application of linguistic approximation transforms abstract fuzzy logic rules into natural language, making the BCI's reasoning accessible to a much wider audience, including clinicians and researchers without deep machine learning expertise.
  • Interpretability as a Tool for Improvement: The paper highlights that an interpretable BCI isn't just for understanding; it can be actively used to verify the learned knowledge, identify inconsistencies with existing literature, and even suggest ways to improve the BCI's performance.

Key Results

  • Interpretable Rules: The algorithm successfully generated "if-then" rules that describe the relationship between brain activity (power in specific regions and frequency bands) and mental states (left vs. right hand intention).
  • Consistency with Literature: The automatically extracted linguistic rules for hand intention prediction were found to be consistent with existing neurophysiological literature, particularly regarding event-related desynchronization in the motor cortex contralateral to the intended hand movement.
  • Performance: The BCI using the raw fuzzy rules achieved an accuracy of 85% on the test set, which was slightly better than the competition winners (84%).
  • Ease of Understanding: The linguistic rules were found to be significantly easier to read and understand than the raw fuzzy rules, making them accessible to non-experts like neurophysiologists.
  • Potential for Improvement: In a separate evaluation on a visual spatial attention task, inspection of the linguistic rules by a neurophysiologist revealed that some rules and features were inconsistent with the literature. Removing these led to a slight increase in classification performance, demonstrating how interpretability can facilitate BCI refinement.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while fuzzy sets are easier to interpret than black-box models, they are still not "naturally easy to understand" for everyone. Linguistic approximation is a step towards this, but further simplification might be needed.
  • The evaluation was performed on a limited number of subjects and specific tasks.
  • Future Work:
    • Conducting additional evaluations on more subjects and different types of neurophysiological signals.
  • Improving the method by incorporating temporal information (the time course of brain activity).
  • Exploring new ways to represent the learned rules to make the BCI even more intuitively interpretable.

Summary updated

Abstract

Most Brain-Computer Interfaces (BCI) are based on machine learning and behave like black boxes, i.e., they cannot be interpreted. However, designing interpretable BCI would enable to discuss, verify or improve what the BCI has automatically learnt from brain signals, or possibly gain new insights about the brain. In this paper, we present an algorithm to design a fully interpretable BCI. It can explain what power in which brain regions and frequency bands corresponds to which mental state, using “if-then ” rules expressed with simple words. Evaluations showed that this algorithm led to a truly interpretable BCI as the automatically derived rules were consistent with the literature. They also showed that we can actually verify and correct what an interpretable BCI has learnt so as to further improve it.Towards a Fully Interpretable EEG-based BCI System 2
BibTeX
@misc{lotte2010towards,
  author = {Fabien Lotte and Anatole Lécuyer and Cuntai Guan},
  title = {{Towards a Fully Interpretable EEG-based BCI System}},
  year = {2010},
}
Figure