When should MI-BCI feature optimization include prior knowledge, and which one?

Camille Benaroch, Maria Sayu Yamamoto, Aline Roc, Pauline Dreyer, Camille Jeunet, Fabien LotteBrain-Computer 2022 · DOI

TL;DR

This paper investigates whether incorporating prior neurophysiological knowledge into data-driven feature selection for Motor Imagery Brain-Computer Interfaces (MI-BCIs) can improve performance. The authors found that while a direct causal link wasn't definitively proven, a constrained approach to feature selection, guided by prior knowledge, showed promise in enhancing user performance, particularly for individuals with distinct or indistinct EEG patterns.

What Problem Does This Paper Address?

Motor Imagery Brain-Computer Interfaces (MI-BCIs) allow users to control devices by imagining movements, which alters brain activity. A crucial step in MI-BCIs is extracting relevant features from brain signals (like EEG) to decode these imagined movements. Current methods often rely heavily on data-driven approaches, meaning they learn features solely from the recorded EEG data with minimal prior knowledge about how the brain works. This raises a key question: could incorporating neurophysiological knowledge into the feature selection process lead to better and more reliable MI-BCI performance? Specifically, the authors want to know when and which prior knowledge should be used.

Methods

The core of this research revolves around the Most Discriminant Frequency Band (MDFB), a feature selection method that identifies the frequency range in EEG signals most correlated with imagined movements.

  1. Correlation Analysis: The researchers first examined the relationship between the characteristics of the MDFB selected by a standard, data-driven algorithm (the unconstrained algorithm) and the user's online MI-BCI performance (classification accuracy). They looked at the MDFB's mean frequency and width.
  2. Constrained Algorithm: To investigate causality, they developed a constrained version of the MDFB selection algorithm. This version incorporated prior knowledge derived from the initial correlation analysis, specifically enforcing certain MDFB characteristics (e.g., a minimum width and a specific frequency range) that were associated with higher performance.
  3. Machine Learning for Prediction and Selection: To further understand the relationship between MDFB characteristics and performance, they used machine learning:
  4. LASSO Regression: To predict MI-BCI performance based on MDFB characteristics.
  5. Decision Trees: To automatically select the optimal algorithm (constrained or unconstrained) for each individual user.

Datasets

The study utilized data from two experiments:

  • Dataset A: This dataset was used for initial correlation analyses and for creating a control group. It contained data from 55 subjects.
  • Dataset B: This dataset was specifically collected for the online experiment comparing the constrained and unconstrained algorithms. It included 21 participants.
  • Dataset A*: A subset of Dataset A, matched to participants in Dataset B based on their MDFB characteristics, served as the control group for the direct comparison of the constrained vs. unconstrained algorithms.

Key Insights and Ideas

  • MDFB Characteristics Matter: The specific frequency band chosen by the algorithm isn't arbitrary; its characteristics (like how wide it is and its central frequency) are correlated with how well a user can control the BCI.
  • Prior Knowledge Can Help (Sometimes): While not a universal performance booster, the constrained algorithm showed potential to improve performance for certain user groups, suggesting that guiding the feature selection process with prior knowledge can be beneficial.
  • Individualization is Key: The study highlights that a "one-size-fits-all" approach to feature selection might not be optimal. Machine learning models can help identify which algorithm (constrained or unconstrained) is best suited for an individual user.
  • The "Why" Behind Performance: By modeling the MDFB selection process with Gaussian functions, the researchers identified parameters that predict performance, offering insights into why some users perform better than others.

Key Results

  • Correlation: A significant positive correlation was found between MDFB width and MI-BCI performance, and a significant negative correlation between MDFB mean frequency and performance.
  • Constrained vs. Unconstrained: While the average online performance for the constrained algorithm was slightly higher (64.7% vs. 62.8%), this difference was not statistically significant. However, the constrained algorithm improved performance for 78% of subjects.
  • Predictive Models: Machine learning models could predict MI-BCI performance better than chance. Key predictors included MDFB width, a parameter representing the "strength" of the frequency band's correlation signal, and the mean frequency of that band.
  • Algorithm Selection: A decision tree model could suggest the most suitable algorithm for a subject. Subjects with very weak or very strong EEG signals (characterized by the "a" parameter) seemed to benefit more from the constrained algorithm.

Limitations and Future Work

  • Statistical Significance: The direct comparison between the constrained and unconstrained algorithms did not reach statistical significance, suggesting that while promising, the improvement wasn't consistently dramatic across all users.
  • Thresholds: The thresholds used to define "high performance" MDFB characteristics were set visually, which could be refined with larger datasets and more sophisticated clustering methods.
  • BCI Illiteracy: A significant portion of participants (52%) performed below chance level, a phenomenon known as BCI illiteracy. The study acknowledges that this could be due to various factors (e.g., short training, mental state, suboptimal training protocols) and suggests future work could focus on classifying users as BCI-literate or -illiterate rather than just predicting performance.
  • Generalizability: The findings are specific to the MDFB algorithm. Future work should explore similar approaches for other feature extraction methods (e.g., Common Spatial Pattern, spatial filters) and classifiers.
  • Causality: While the constrained approach aimed to explore causality, definitive proof remains elusive. Future research could further investigate causal links.

Summary updated

Abstract

Motor imagery-based brain–computer interfaces (MI-BCIs) rely on interactions between humans and machines. The (learning) characteristics of both components are key to understand and improve performances. Data-driven methods are often used to select/extract features with very little neurophysiological prior. Should they include prior knowledge and which one? We study the relationship between MI-BCI performances and characteristics of a commonly used subject-specific Most Discriminant Frequency Band (MDFB). Our results showed a correlation between the selected MDFB characteristics (mean and width) and performances. While a causal link could not be determined, online MI-BCI performances obtained using a constrained algorithm --enforcing characteristics associated to high performances– seemed higher than with an unconstrained algorithm. Finally, we used machine learning to 1) predict MI-BCI performances from MDFB characteristics and 2) select automatically the optimal algorithm. This way, the constrained algorithm could improve performances for subjects with either clearly distinct or no distinct EEG patterns.
BibTeX
@inproceedings{benaroch2022when,
  author = {Camille Benaroch and Maria Sayu Yamamoto and Aline Roc and Pauline Dreyer and Camille Jeunet and Fabien Lotte},
  title = {{When should MI-BCI feature optimization include prior knowledge, and which one?}},
  booktitle = {Brain-Computer},
  year = {2022},
  doi = {10.1080/2326263x.2022.2033073},
}
Figure