PaperlayerView exact revision on arXiv

Source-linked AI summary

FBCNet: A Multi-view Convolutional Neural Network for Brain-Computer Interface

Ravikiran Mane, Effie Chew, Karen Chua, Kai Keng Ang, Neethu Robinson, A. P. Vinod, Seong-Whan Lee, Cuntai Guan

arXiv:2104.01233v1cs.OHcs.AIcs.LGeess.SP

TL;DR

MI EEG-BCI decoding must handle noisy, high-dimensional signals and limited training data, especially in underexplored stroke populations. FBCNet addresses this with a neurophysiologically inspired filter-bank CNN and Variance layer, achieving the best reported performance across four datasets, including 76.20% 4-class accuracy on BCIC-IV-2a.

  • Problem

    Noisy high-dimensional EEG, variability, and scarce training data challenge MI classification, while deep-learning evaluation remains largely limited to healthy participants.

  • Method

    FBCNet combines multi-view spectral filtering, CNN-based spatial filtering, and a Variance layer that extracts compact temporal information.

  • Results

    FBCNet achieved the best classification performance across all evaluated datasets, including 76.20% 4-class accuracy on BCIC-IV-2a in hold-out analysis.

  • Takeaways & Limitations

    FBCNet’s neurophysiological design achieved significantly better accuracies across four MI datasets, including two chronic-stroke datasets, while learning neurophysiologically relevant EEG features.

Abstract

from arXiv · show

Lack of adequate training samples and noisy high-dimensional features are key challenges faced by Motor Imagery (MI) decoding algorithms for electroencephalogram (EEG) based Brain-Computer Interface (BCI). To address these challenges, inspired from neuro-physiological signatures of MI, this paper proposes a novel Filter-Bank Convolutional Network (FBCNet) for MI classification. FBCNet employs a multi-view data representation followed by spatial filtering to extract spectro-spatially discriminative features. This multistage approach enables efficient training of the network even when limited training data is available. More significantly, in FBCNet, we propose a novel Variance layer that effectively aggregates the EEG time-domain information. With this design, we compare FBCNet with state-of-the-art (SOTA) BCI algorithm on four MI datasets: The BCI competition IV dataset 2a (BCIC-IV-2a), the OpenBMI dataset, and two large datasets from chronic stroke patients. The results show that, by achieving 76.20% 4-class classification accuracy, FBCNet sets a new SOTA for BCIC-IV-2a dataset. On the other three datasets, FBCNet yields up to 8% higher binary classification accuracies. Additionally, using explainable AI techniques we present one of the first reports about the differences in discriminative EEG features between healthy subjects and stroke patients. Also, the FBCNet source code is available at https://github.com/ravikiran-mane/FBCNet.

I. INTRODUCTION

MI EEG-BCI decoding is hindered by noisy, high-dimensional signals, variability, and limited training data. FBCNet addresses these constraints by combining neurophysiological priors with deep learning and evaluates performance across healthy and stroke populations.

  • Research challenges: Noisy high-dimensional data, inter-trial variance, intra-class variability, and scarce training data make MI classification especially difficult.These challenges motivate both classical feature-engineering approaches and deep learning methods.
  • Existing approaches: Classical methods exploit neurophysiologically grounded sensory-motor rhythms but remain susceptible to intra-trial variance and handcrafted-feature dependence.FBCSP and Riemannian methods perform well with limited noisy EEG, yet retain these constraints.
  • Existing approaches: Deep learning offers end-to-end feature learning, but limited data and high feature dimensionality can cause overfitting and only marginal improvements.CNNs are attractive because they learn local connectivity patterns directly from EEG.
  • Clinical motivation: Deep-learning evaluation in MI classification has largely focused on healthy participants, despite altered brain dynamics and motor control after stroke.The paper motivates testing whether these methods improve MI-BCI performance for stroke patients.
  • Proposed contribution: FBCNet is a compact, neurophysiologically inspired end-to-end CNN that combines spectral filtering, CNN spatial filtering, and a Variance layer for temporal information.The architecture is designed for subject-specific MI classification with limited data and reduced parameters.
  • Study scope: The study compares classical and deep learning approaches for MI decoding in large healthy and chronic-stroke populations and examines discriminative EEG differences.The contributions include reporting deep-learning effectiveness in stroke patients and comparing healthy and stroke data.

II. RELATED WORKS

FBCNet follows a filter-bank, spatial-filtering, variance-based pipeline that embeds MI neurophysiology into an end-to-end CNN. Its multi-view representation localizes EEG spectrally before compact temporal classification.

  • Comparison methods: FBCNet is compared directly with FBCSP because both share a similar design philosophy, while Deep ConvNet and EEGNet provide deep-learning baselines.The evaluation also includes comparisons with recent architectures using reported BCIC-IV-2a accuracies.
  • Multi-view representation: FBCNet first creates multiple EEG views by applying narrow-band filters to the raw signal, localizing information across frequency bands.The design targets MI information concentrated in the mu and beta bands.
  • Spatial transformation learning: A depthwise convolution learns spatially discriminative patterns separately for each spectrally localized view.This stage follows deterministic spectral localization and performs spatial localization of discriminative EEG features.
  • Temporal extraction and classification: The Variance layer extracts temporal information into a compact representation, after which a fully connected layer assigns class labels.The pipeline uses variance-based aggregation for dimensionality reduction before classification.
  • Architecture notation: In the architecture notation, C and T denote EEG channels and time points, while Nb, m, and Nc denote frequency bands, filters per band, and output classes.These dimensions describe the inputs and configurable network components shown in Fig. 1.
  • Spectral filtering: The filter bank uses nine non-overlapping 4-Hz bands spanning 4–40 Hz with Chebyshev Type II filtering.The stated transition bandwidth is 2 Hz and stopband ripple is -30 dB.

2) Spatial Localization by CNN:

FBCNet uses depthwise convolution for frequency-band-specific spatial filtering and a Variance layer for temporally discriminative EEG features. The Variance layer emphasizes deviations from the signal mean and reduces temporal dimensionality before classification.

  • Spatial filtering: Depthwise convolution learns spatially discriminative patterns separately for each frequency band, with m controlling the number of spatial filters per band.
  • Temporal feature extraction: Variance characterizes EEG time series through spectral-power-related temporal variation, offering an alternative to max or average pooling.
  • Temporal feature extraction: During backpropagation, the Variance layer assigns a higher proportion of incoming gradient to signal points farther from the mean, aligning with ERD and ERS signatures.
  • Temporal feature extraction: The Variance layer computes temporal variance over non-overlapping windows after the SCB output.
  • Temporal feature extraction: Using window length w = 1s reduces features from (m×N_b×T) to (m×N_b×T/w), while different window lengths are evaluated in ablation analysis.
  • Classification: Log-activated Variance-layer features are passed to a linear fully connected layer and then a softmax layer for class probabilities.

4) Classification:

The evaluation spans four diverse EEG-MI datasets covering healthy participants and stroke patients. These include one 4-class benchmark and three binary MI datasets, with two datasets focused on post-stroke rehabilitation.

  • Evaluation datasets: FBCNet was evaluated on four EEG-MI datasets to address extreme inter-subject variability in algorithm performance.
  • Healthy-participant datasets: The BCIC-IV-2A dataset contains 4-class MI data, while OpenBMI contains 2-class MI data.
  • Stroke datasets: Stroke Data A and Stroke Data B contain 2-class MI-versus-rest data from large stroke datasets.
  • Healthy-participant datasets: BCIC-IV-2A and OpenBMI contain EEG data from healthy people and are publicly available benchmarks for state-of-the-art classification algorithms.
  • Stroke datasets: The paper includes the two stroke datasets because post-stroke motor rehabilitation is a focus of the study.

C. Experiments

The experiments combine 10-fold cross-validation with subject-specific inter-session hold-out analysis and compare FBCNet with classical and CNN-based algorithms. Cross-validation excludes inter-session data to avoid confounding from inter-session variability.

  • Evaluation design: FBCNet was evaluated using both cross-validation and hold-out analyses.
  • Cross-validation: The cross-validation analysis used 10 folds, with nine folds for training and one fold for testing.
  • Cross-validation: Cross-validation used complete stroke-dataset data and session 1 data from BCIC-IV-2A and OpenBMI.
  • Cross-validation: Inter-session data was excluded from cross-validation to avoid the known confounding influence of inter-session variability.
  • Hold-out analysis: The hold-out analysis trained on each subject’s session 1 data and tested the resulting model on session 2 data for BCIC-IV-2A and OpenBMI.
  • Baselines: FBCNet was compared with FBCSP-SVM, Deep ConvNet, and EEGNet-8,2 using settings recommended by their respective authors.

E. Interpretability and Visualizations

Across datasets, FBCNet generally achieved the strongest classification performance, remained robust with limited training data, and its Variance layer produced the best temporal feature extraction results.

  • Overall performance: FBCNet achieved the best classification performance across datasets in both cross-validation and holdout analyses, including 76.20% 4-class accuracy on BCIC-IV-2A.The maximum improvement over FBCSP-SVM was 10% in OpenBMI cross-validation, with statistically significant gains in most datasets.
  • Subject-level performance: FBCNet matched the best-performing method for most subjects, outperforming deep architectures among weaker subjects and classical methods among stronger subjects.For the highest-accuracy subjects, FBCNet exceeded FBCSP-SVM; for the lowest-accuracy subjects, it also achieved the highest average accuracy.
  • Subject-level performance: FBCNet achieved accuracy above 70% for the most subjects across datasets, including 28% more stroke patients than FBCSP-SVM.The 70% threshold was analyzed as a practical usability criterion for binary classification.
  • Training-set sensitivity: FBCNet retained relatively strong performance with small training sets while matching deep-learning accuracy when ample training data were available.Deep ConvNet and EEGNet-8,2 were highly sensitive to reduced training data, whereas FBCSP-SVM was less affected but weaker with complete data.
  • Temporal feature extraction: The Variance layer produced the highest classification accuracies across all analyses, significantly outperforming Average and Max temporal extraction layers.The reported differences between Variance and the alternative layers were statistically significant across datasets.
  • Architecture sensitivity: Increasing spatial filters and reducing variance-window size initially improved accuracy, but excessive filter counts or overly small windows reduced accuracy.The number of spatial filters and window length jointly affect model capacity and computational complexity.

B. Interpretability and visualizations

Relevance analysis identified concentrated motor-area, mid-frequency patterns in healthy subjects but more diffuse and variable frequency-channel relevance in stroke patients.

  • Analysis design: The analysis compared subject-averaged channel-frequency relevance patterns and individual-subject relevance distributions between healthy and stroke groups.The results were obtained using group-level relevance analysis of FBCNet inputs.
  • Group-level relevance: Healthy subjects showed strongest relevance in the 8–16 Hz range, concentrated around the C3 and C4 motor-area channels.The 8–12 Hz and 12–16 Hz bands together accounted for 34% of subject-averaged input relevance.
  • Group-level relevance: Stroke patients displayed greater inter-subject variability in relevance patterns than healthy subjects.Across stroke patients, the most relevant frequency range differed substantially between individuals, unlike the more consistent 8–12 Hz pattern in healthy subjects.

V. DISCUSSION

FBCNet combines neurophysiological priors with deep learning for MI classification and was evaluated across large healthy and stroke cohorts alongside interpretability analysis.

  • Discussion: FBCNet is a neurophysiologically inspired, end-to-end CNN designed to learn generalizable discriminative features with limited data.The evaluation included 63 healthy subjects and 71 chronic stroke patients.
  • Discussion: The study compared classical and deep-learning approaches across healthy and stroke EEG-MI datasets and examined differences in learned relevance patterns.The paper presents the comparison for a large population spanning both groups.

A. Performance of classical machine learning methods and state-of-the-art deep learning architectures

Baseline deep-learning methods did not substantially outperform classical methods, motivating FBCNet’s hybrid design, which achieved strong performance across varied subjects and data regimes.

  • Baseline comparison: Classical FBCSP and baseline deep networks showed no significant subject-averaged accuracy difference in all but one analysis.The authors report that deep learning had not yet produced substantial improvements in EEG classification.
  • Baseline comparison: Reduced training data may have affected deep-learning architectures more than FBCSP, especially in the stroke datasets with 20% fewer training trials than OpenBMI.The authors present this as a possible explanation for dataset-level baseline differences.
  • Subject variability: FBCSP produced extreme subject-level accuracies, whereas baseline deep networks more often produced average accuracies, consistent with inter-subject EEG distribution variability.The authors identify inter-subject variability as one possible explanation for these patterns.
  • FBCNet design: FBCNet’s hybrid architecture incorporates neurophysiological knowledge into deep learning to balance model capacity and complexity for small, noisy EEG datasets.This design was intended to reduce susceptibility to limited and noisy data while retaining feature-learning capacity.
  • FBCNet design: The Variance layer was motivated by the relationship between filtered-signal variance and spectral power, a class-discriminative property of MI EEG.Ablation results confirmed the Variance layer’s advantage over alternative temporal consolidation operations.
  • FBCNet design: FBCNet’s spatial-filter count and variance-window length control representation capacity and computational complexity, with m between 16 and 32 and w between 0.5s and 1s often performing best.More complex 4-class BCIC-IV-2A classification benefited from larger m values.

E. Interpretability Analysis

FBCNet’s interpretability analysis links its decisions to neurophysiologically relevant EEG features and reveals different discriminative patterns in healthy subjects and stroke patients.

  • Healthy subjects: FBCNet assigned healthy-subject relevance to C3 and C4 in the 8–16 Hz and 24–32 Hz ranges, consistent with motor-imagery signatures.These patterns also generalized to unseen test data and accompanied higher classification accuracies.
  • Stroke patients: Stroke patients showed more spatially and spectrally diffuse relevance patterns with substantial inter-subject variability.The authors associate these differences with stroke-induced, subject-specific brain modifications.
  • Stroke patients: The analysis suggests that stroke-related brain damage may involve compensatory recruitment of non-motor areas during motor imagery.The authors note that confirming this phenomenon requires more data with distinct motor-imagery tasks from stroke patients.
  • Interpretability implications: FBCNet combines neurophysiological priors with deep-learning feature learning to obtain generalizable discriminative features despite limited training data.Its design uses a hybrid approach that constrains learning toward motor-imagery signatures while retaining deep-learning capacity.

S1 FBCNet architecture

The supplementary materials describe FBCNet’s implementation context and the four MI datasets used for evaluation, including two post-stroke datasets.

  • Datasets: FBCNet was evaluated on four datasets: BCIC-IV-2A, OpenBMI, Stroke Data A, and Stroke Data B.The benchmark includes one 4-class dataset and three 2-class datasets.
  • Datasets: BCIC-IV-2A contains 4-class MI data from 9 subjects across two sessions, recorded with 22 electrodes at 250 Hz.Classes are left hand, right hand, feet, and tongue imagery; each trial lasts 4 seconds.
  • Datasets: OpenBMI contains two-session left- versus right-hand MI data from 54 healthy subjects, originally recorded with 62 electrodes at 1000 Hz.The analysis selects 20 channels in the motor region and uses 4-second trials.
  • Datasets: Stroke Data A contains stroke-paralyzed-hand MI versus rest recordings from 37 patients, using 27 channels sampled at 250 Hz.Each patient contributes 80 trials of each class, with 4-second trials and hardware filtering from 0.5–40 Hz.
  • Datasets: Stroke Data B contains the same two classes from 34 stroke patients, with 160 trials per patient and the same 27-channel, 250 Hz protocol.The analysis uses the entire 27-channel, 4-second data for classification.

S3 Implementation of Baseline Classification methods

The baseline implementation applies filter-bank CSP-SVM and established deep-learning architectures, with dataset-specific adjustments and matched filtering for fair comparison.

  • FBCSP-SVM: FBCSP-SVM decomposes EEG into nine 4-Hz Chebyshev Type II bands spanning 4–40 Hz, then applies CSP and log-variance features.Four most discriminative CSP filters are extracted from each band for binary classification.
  • FBCSP-SVM: For 4-class BCIC-IV-2A, FBCSP-SVM uses four one-versus-rest binary classifiers and assigns each trial the class with the maximum score.This produces 16 CSP filters per band across the four OVR models.
  • Comparison protocol: FBCSP and FBCNet use the same filter bank to ensure a fair comparison based on the multi-view filtered EEG representation.The shared preprocessing isolates architectural differences between the methods.
  • Deep-learning baselines: Deep ConvNet and EEGNet-8,2 were implemented from their original descriptions, with EEGNet temporal dimensions doubled for 250-Hz data.The adjustment accounts for the original EEGNet design targeting 128-Hz sampling.
  • Results reporting: Supplementary tables report mean±std subject-specific accuracies, individual-subject results, and stroke-patient accuracies for reproducibility and comparison.These results are provided in Tables S2–S5.
Loading 2104.01233v1…