[Paper Review] Functional connectivity ensemble method to enhance BCI performance (FUCONE)

Marie‐Constance Corsi, Sylvain Chevallier+2 more|arXiv (Cornell University)|Nov 4, 2021

EEG and Brain-Computer Interfaces79 references19 citations

TL;DR

FUCONE proposes a novel ensemble learning framework that combines multiple functional connectivity estimators with Riemannian geometry-based classification to enhance motor-imagery BCI performance. By training a Riemannian classifier on each FC estimator and ensembling their decisions, FUCONE achieves significant performance gains across diverse datasets, outperforming state-of-the-art methods due to improved feature space diversity and robustness to inter- and intra-subject variability.

ABSTRACT

Functional connectivity is a key approach to investigate oscillatory activities of the brain that provides important insights on the underlying dynamic of neuronal interactions and that is mostly applied for brain activity analysis. Building on the advances in information geometry for brain-computer interface, we propose a novel framework that combines functional connectivity estimators and covariance-based pipelines to classify mental states, such as motor imagery. A Riemannian classifier is trained for each estimator and an ensemble classifier combines the decisions in each feature space. A thorough assessment of the functional connectivity estimators is provided and the best performing pipeline, called FUCONE, is evaluated on different conditions and datasets. Using a meta-analysis to aggregate results across datasets, FUCONE performed significantly better than all state-of-the-art methods. The performance gain is mostly imputable to the improved diversity of the feature spaces, increasing the robustness of the ensemble classifier with respect to the inter- and intra-subject variability.

Motivation & Objective

• Address the 'BCI inefficiency' problem, where up to 30% of users fail to master BCI control due to high inter- and intra-subject variability.
• Explore whether functional connectivity (FC) estimators can provide complementary, robust features for classifying mental states in motor-imagery-based BCIs.
• Develop and optimize an ensemble learning framework that combines multiple FC estimators and Riemannian classifiers to improve classification accuracy.
• Ensure the method's robustness across diverse datasets with varying channel counts, task designs, and session configurations.
• Provide insights into the neural mechanisms underlying BCI performance by analyzing the contribution of different FC estimators.

Proposed method

• Apply a variety of functional connectivity estimators—including spectral coherence, phase-locking value, phase-lagged index, and amplitude envelope coupling—on EEG signals to extract connectivity features.
• Transform each FC feature matrix into a symmetric positive-definite (SPD) matrix, suitable for Riemannian geometry-based classification.
• Train a Riemannian classifier (e.g., Riemannian MDM or K-Nearest Neighbor) independently on each FC estimator’s feature space.
• Combine the individual classifier decisions using an ensemble averaging strategy to produce a final prediction.
• Optimize key parameters such as frequency bands and FC estimators using a reference dataset before evaluating on eight independent MI datasets.
• Apply dimensionality reduction to identify a minimal, high-performing electrode subset, improving clinical applicability.

Experimental results

Research questions

• RQ1 Can combining multiple functional connectivity estimators through ensemble learning significantly improve BCI classification performance compared to single-estimator approaches?
• RQ2 Which functional connectivity estimators contribute most effectively to Riemannian-based classification of motor-imagery tasks?
• RQ3 How robust is the FUCONE framework across diverse datasets with varying numbers of EEG channels, task paradigms, and session configurations?
• RQ4 Does the ensemble of diverse FC-based feature spaces reduce sensitivity to inter- and intra-subject variability in BCI performance?
• RQ5 Can a dimensionality reduction step identify a minimal, clinically relevant electrode subset that maintains high classification accuracy?

Key findings

• FUCONE significantly outperformed all state-of-the-art methods in a meta-analysis across eight diverse motor-imagery datasets, demonstrating superior generalization and robustness.
• The ensemble approach achieved higher accuracy than both traditional CSP-based and single Riemannian covariance-based classifiers, with performance gains attributed to enhanced diversity in feature spaces.
• The method remained effective under both within-session and cross-session evaluation protocols, indicating strong adaptability to real-world BCI deployment scenarios.
• A dimensionality reduction step identified a subset of electrodes—primarily over sensorimotor and associative cortical areas—as sufficient for high accuracy, supporting clinical feasibility.
• The analysis revealed that phase-based estimators like wPLI2-d and PLI contributed most significantly to the ensemble’s performance, highlighting their relevance in capturing task-related brain network dynamics.
• The processing delay for 1-second epochs was approximately 100 ms, confirming feasibility for online BCI implementation with overlapping epochs.