ICLabel: An automated electroencephalographic independent component classifier, dataset, and website (1901.07915v2)

Abstract: The electroencephalogram (EEG) provides a non-invasive, minimally restrictive, and relatively low cost measure of mesoscale brain dynamics with high temporal resolution. Although signals recorded in parallel by multiple, near-adjacent EEG scalp electrode channels are highly-correlated and combine signals from many different sources, biological and non-biological, independent component analysis (ICA) has been shown to isolate the various source generator processes underlying those recordings. Independent components (IC) found by ICA decomposition can be manually inspected, selected, and interpreted, but doing so requires both time and practice as ICs have no particular order or intrinsic interpretations and therefore require further study of their properties. Alternatively, sufficiently-accurate automated IC classifiers can be used to classify ICs into broad source categories, speeding the analysis of EEG studies with many subjects and enabling the use of ICA decomposition in near-real-time applications. While many such classifiers have been proposed recently, this work presents the ICLabel project comprised of (1) an IC dataset containing spatiotemporal measures for over 200,000 ICs from more than 6,000 EEG recordings, (2) a website for collecting crowdsourced IC labels and educating EEG researchers and practitioners about IC interpretation, and (3) the automated ICLabel classifier. The classifier improves upon existing methods in two ways: by improving the accuracy of the computed label estimates and by enhancing its computational efficiency. The ICLabel classifier outperforms or performs comparably to the previous best publicly available method for all measured IC categories while computing those labels ten times faster than that classifier as shown in a rigorous comparison against all other publicly available EEG IC classifiers.

• The paper introduces ICLabel, which automates EEG IC classification with a CNN-based model and a large, crowdsourced dataset.
• It integrates scalp topographies, PSDs, and ACFs to achieve superior balanced accuracy and ten times faster performance than existing methods.
• The project offers an open MATLAB tool and educational website to support efficient, scalable EEG analysis in research and BCI applications.

ICLabel: An Automated Electroencephalographic Independent Component Classifier, Dataset, and Website

The described paper presents ICLabel, a comprehensive project designed to improve the classification of EEG independent components (ICs). Recognizing the complexity and time requirements of manual IC inspection and interpretation following Independent Component Analysis (ICA), the authors developed ICLabel to automate and enhance the IC classification process. The ICLabel project includes a dataset of labeled ICs, an online platform for crowdsourcing IC labels, and a robust classifier that drives efficiency and accuracy in IC categorization.

Overview of ICLabel Components

1. ICLabel Dataset: The dataset underpins the ICLabel classifier and comprises spatiotemporal measures for over 200,000 ICs from more than 6,000 EEG recordings. These records have undergone common average reference conversion and include labels for over 6,000 ICs derived through crowdsourcing.
2. ICLabel Website: The platform (\url{https://iclabel.ucsd.edu/tutorial}) serves dual purposes as an educational tool and a crowdsourcing mechanism. It facilitates IC labeling by a diverse range of contributors, collecting over 34,000 labels from more than 250 users.
3. ICLabel Classifier: The automated classifier, freely available for MATLAB, achieves superior accuracy and computational efficiency. Notably, it outperforms existing publicly available IC classifiers by a significant margin in speed, achieving ten times faster calculations.

Performance and Methodology

The ICLabel classifier operates by taking input features such as scalp topographies, power spectral densities (PSDs), and autocorrelation functions (ACFs) of ICs, which are then processed by a convolutional neural network (CNN). Addressing the inherent class imbalances in EEG data, the architecture benefits from detailed feature extraction processes and balanced accuracy frameworks.

Candidate Classifier Evaluation:

• CNN and GAN Architectures: Multiple candidate classifiers employing CNNs and generative adversarial networks (GANs) were evaluated. The GAN-based approaches did not demonstrate superior performance compared to CNNs. The Weighted CNN (wCNN) using topographies, PSDs, and ACFs (wCNNAC) emerged as the best performer across diverse classification metrics.
• Cross-validation: A ten-fold cross-validation approach was employed to refine model performance and generalization, revealing that wCNNAC provided the most balanced accuracy and lower cross-entropy across IC classes.

Results

Comparative Analysis:

• Balanced Accuracy and Cross Entropy: ICLabel exhibited balanced accuracy figures of 0.841 for two-class and notable improvements in the five- and seven-class comparisons. Cross-entropy values were consistently low, indicating high-confidence classification.
• Speed of Classification: The median classification time for ICs by ICLabel was approximately 170 ms, demonstrating a substantial improvement over competing methods like IC_MARC, which required closer to 1.8 seconds per IC.

Practical and Theoretical Implications

Practical Implications:

ICLabel’s enhanced classification rate has essential applications for large-scale EEG studies and real-time EEG analysis, enabling rapid, reliable IC categorization. This acceleration is particularly beneficial for brain-computer interface (BCI) systems and other near-real-time EEG applications.

Theoretical Contributions:

ICLabel’s development and evaluation enhance our understanding of automatic IC classification’s feasibility and reliability. The use of diverse spatiotemporal features and advanced neural network architectures provides a robust foundation for future EEG decomposition and artifact removal methods.

Conclusion

ICLabel represents a significant advancement in the automated classification of EEG independent components, striking a balance between accuracy and computational efficiency. This project not only provides a scalable solution to IC categorization but also lays the groundwork for further advancements in real-time neural data analysis and education around EEG data interpretation. The openly available ICLabel classifier and dataset invite continuous community engagement, promising sustained improvements and adaptations in the evolving landscape of EEG research. The tools and methodologies propagated by ICLabel are poised to significantly impact both academic research and practical implementations in neuroengineering.