Source-linked AI summary
EEG-Inception: An Accurate and Robust End-to-End Neural Network for EEG-based Motor Imagery Classification
Ce Zhang, Young-Keun Kim, Azim Eskandarian
TL;DR
EEG-based motor-imagery classification must handle non-stationary, noisy signals, small datasets, and subject variability. The paper proposes EEG-Inception, an end-to-end CNN with EEG data augmentation, and reports strong benchmark accuracy with preliminary subject-independent results.
Problem
EEG-MI classification is limited by non-stationary, noisy signals, small datasets, and subject-dependent models that require training data for new subjects.
Method
EEG-Inception modifies an Inception-Time-style CNN with inception and residual modules and adds EEG signal augmentation to address limited training data.
Results
EEG-Inception reports 88.58% accuracy for binary classes and 88.39% for four classes, with subject-independent accuracies of 77.5% and 65.9%, respectively.
Takeaways & Limitations
The reported results support EEG-Inception’s potential for robust EEG-MI classification and further subject-independent application.
Takeaways & Limitations
Subject-independent results are preliminary because the subject-dependent model was applied without further fine-tuning, and the four-class dataset was not large enough for this analysis.
Abstract
from arXiv · showhide
Classification of EEG-based motor imagery (MI) is a crucial non-invasive application in brain-computer interface (BCI) research. This paper proposes a novel convolutional neural network (CNN) architecture for accurate and robust EEG-based MI classification that outperforms the state-of-the-art methods. The proposed CNN model, namely EEG-Inception, is built on the backbone of the Inception-Time network, which showed to be highly efficient and accurate for time-series classification. Also, the proposed network is an end-to-end classification, as it takes the raw EEG signals as the input and does not require complex EEG signal-preprocessing. Furthermore, this paper proposes a novel data augmentation method for EEG signals to enhance the accuracy, at least by 3%, and reduce overfitting with limited BCI datasets. The proposed model outperforms all the state-of-the-art methods by achieving the average accuracy of 88.4% and 88.6% on the 2008 BCI Competition IV 2a (four-classes) and 2b datasets (binary-classes), respectively. Furthermore, it takes less than 0.025 seconds to test a sample suitable for real-time processing. Moreover, the classification standard deviation for nine different subjects achieves the lowest value of 5.5 for the 2b dataset and 7.1 for the 2a dataset, which validates that the proposed method is highly robust. From the experiment results, it can be inferred that the EEG-Inception network exhibits a strong potential as a subject-independent classifier for EEG-based MI tasks.
1. Introduction
EEG-based MI classification is challenging because signals are non-stationary, noisy, and limited by small datasets and subject variability. EEG-Inception addresses these issues with a CNN, data augmentation, and a focus on accuracy and robustness.
- EEG-based MI signals are non-stationary and have low signal-to-noise ratios because of artifacts and noise.
- 12 existing neural-network issues include improving accuracy, limited EEG dataset sizes, and dependence on subject-specific training data.
- EEG-Inception combines a new CNN with a data-augmentation method intended to improve accuracy and robustness to subject-dependency.
- 3% is the reported average-accuracy increase from the proposed EEG signal augmentation method.
- 88.6% average accuracy is reported for binary-class datasets, while 88.4% is reported for four-class datasets, outperforming other state-of-the-art methods.
- The paper identifies augmentation, low subject-dependent variability, and subject-independent potential as key contributions.
2. Related Works
Related work contrasts conventional pipelines requiring manual preprocessing and feature extraction with neural networks that learn EEG representations. The paper identifies remaining needs for higher accuracy and effective subject-independent classification.
- Conventional EEG-MI classification uses preprocessing, feature extraction, and classification, with methods including ICA, CCA, filtering, KNN, LDA, and SVM.
- Manual feature extraction limits conventional methods because performance depends heavily on the effectiveness of features derived from existing algorithms.
- CNNs and RNNs are the two popular neural-network structures used for EEG-based MI classification.
- 1D-CNNs extract time-series or frequency-domain features, whereas 2D/3D-CNNs are used to extract features from multiple dimensions simultaneously.
- LSTM-based approaches use temporal correlations in raw EEG signals for MI prediction, including a reported 76.47% average accuracy for one Group I dataset.
- Prior work concludes that EEG-MI accuracy still needs improvement and that effective subject-independent algorithms remain underdeveloped.
2. Proposed Network Architecture
EEG-Inception combines inception and residual modules with a novel EEG noise-augmentation method to extract multiscale features and expand limited training data.
- EEG-Inception architecture: EEG-Inception uses an ensemble of inception and residual modules as its backbone.Inception modules extract features across depth and width, while residual modules address degradation associated with deeper layers.
- EEG-Inception architecture: The network comprises six inception modules and two residual modules, with residual modules inserted after every three inception modules.The residual connection sums the residual-layer output with the inception-module output.
- Inception module design: Initial inception modules expand input depth, whereas intermediate modules reduce data dimension for lower computation load.The initial depth changes from 3 to 12 for binary classes and from 22 to 48 for four classes.
- Inception module design: Parallel 1D convolutions use multiple kernel lengths to extract EEG features at different temporal scales.Binary-class models use kernel sizes 25, 75, and 125; four-class models add 175 and 225.
- Data augmentation: The augmentation method extracts noise candidates with a 100 Hz high-pass Butterworth filter, subtracts them from original signals, and recombines them across trials.Training data are increased threefold for the binary dataset and sixfold for the four-classes dataset.
- Experimental analysis: An ablation study varies convolutional-kernel depth to analyze its effect on neural-network classification performance.The tested convolutional-kernel dimension ranges from six to sixty-four for the binary dataset and continues to a higher range for the four-classes dataset.
3. Dataset and Experiment Protocol
The experiments use BCI Competition IV datasets 2a and 2b, covering four-class and binary motor-imagery tasks across nine subjects. Dataset-specific sessions, channel counts, trial structures, and input modifications define the evaluation protocol.
- Datasets: BCI Competition IV datasets 2a and 2b provide EEG and EOG signals sampled at 250 Hz from nine subjects.Dataset 2a uses four MI classes, whereas dataset 2b uses binary left- and right-hand MI.
- Datasets: Dataset 2a contains 22 EEG channels, while dataset 2b contains three EEG channels.The channel localizations are presented in Figure 6.
- Session paradigms: Dataset 2b includes five sessions, with the first two lacking results feedback and the last three providing feedback.The no-feedback sessions contain 120 trials per subject, with three seconds of motor imagery per trial.
- Session paradigms: Dataset 2a contains two sessions without results feedback and 72 trials per class, totaling 288 trials per subject.Each trial includes a three-second motor-imagery period.
- Input preparation: For dataset 2b, the last three sessions are doubled by splitting each six-second imagery period into two three-second trials, yielding 1200 trials per subject.Rejected trials are removed, and training-to-testing data are allocated at approximately 3:1.
4. Results and Discussions
EEG-Inception is evaluated through training, ablation, augmentation, benchmark, and subject-independent analyses. The reported results show faster convergence, strong accuracy and robustness, low single-sample computation time, and reduced performance in the preliminary subject-independent setting.
- Evaluation design: The evaluation covers training behavior, convolution-depth ablation, augmentation effectiveness, state-of-the-art accuracy comparisons, and preliminary subject-independent classification.Experiments use open-source BCI Competition IV datasets 2a and 2b.
- Ablation study: Increasing convolution-kernel depth linearly increases parameters and model size, while excessive depth can cause overfitting.The selected depths are twelve for dataset 2b and forty-eight for dataset 2a; model sizes are 10.83 and 34.10 megabytes, respectively.
- Data augmentation: 10-20 iterations are sufficient for augmented training to converge, compared with 40-60 iterations without augmentation.The augmentation method also improves average accuracy by 2.8% for binary classes and 3.6% for four classes.
- Classification results: 80% is maintained for the poorest-performing binary-dataset subjects, approximately 12.5% above other state-of-the-art algorithms.The authors associate these subjects with unclear neuronal features and artifact contamination.
- Efficiency and robustness: 0.0187 seconds is required to test one binary-dataset sample, while four-class testing takes 0.0215 seconds and remains suitable for real-time processing.Parallel convolutional layers extract features over different time lengths, supporting computational efficiency.
- Efficiency and robustness: 5.5 is the average subject standard deviation for the binary dataset, while the four-class standard deviation is 7.06.The authors interpret these low values as robustness across subjects and potential for subject-independent learning.
- Classification results: 88.39% is the four-class EEG-Inception accuracy, ranking second among the compared state-of-the-art algorithms.The paper attributes the lower four-class result than HS-CNN to smaller training data and the larger parameter-learning demand of five parallel kernels.
- Subject-independent examination: Subject-independent accuracy falls to 77.5% for binary classes and 65.88% for four classes when the subject-dependent model is applied without fine-tuning.Binary subject-independent augmentation raises AUC by 0.03 over training without augmentation.
5. Conclusions and Future Works
EEG-Inception combines modified Inception Time with EEG data augmentation for motor-imagery classification, reducing overfitting with small datasets. It achieves high accuracy and low subject-wise variability, while subject-independent performance remains a preliminary result requiring further refinement.
- EEG-Inception combines a modified Inception Time network with a novel EEG data augmentation method.The augmentation method is reported to diminish overfitting caused by small data size.
- 88.58% average accuracy was achieved on the binary-classes dataset, with 88.39% on the four-classes dataset.
- 5.5 and 7.1 standard deviations across subjects were obtained for the binary-classes and four-classes datasets, respectively.The paper interprets the low standard deviations as evidence of robust feature extraction across subjects.
- 77.5% and 65.9% classification accuracy were obtained in the preliminary subject-independent analysis for binary-classes and four-classes, respectively.
- Further investigation of layer specificity and generality is planned before fine-tuning EEG-Inception for subject-independent classification.