A2DM: Enhancing EEG Artifact Removal by Fusing Artifact Representation into the Time-Frequency Domain

Artificial Intelligence Neurology Information Science Medicine
Electroencephalography Artifact Representation EEG Denoising Frequency Domain Representation

Academic Background

Electroencephalogram (EEG) is a crucial tool for studying brain activity, widely used in neuroscience, clinical diagnosis, and brain-computer interfaces. However, EEG signals are often contaminated by various artifacts during acquisition, such as electrooculography (EOG) and electromyography (EMG). These artifacts significantly degrade the quality of EEG signals, thereby affecting subsequent analysis and applications. Although some methods have been developed to remove single types of artifacts, existing approaches often underperform when dealing with multiple artifacts simultaneously. Therefore, developing a model capable of uniformly removing multiple types of artifacts has become a significant challenge in current research.

Haoran Li et al. addressed this issue by proposing an artifact-aware EEG denoising model called A2DM (Artifact-Aware Denoising Model). This model effectively removes multiple artifacts by fusing artifact representations into the time-frequency domain, significantly improving the quality of EEG signals.

Source of the Paper

The paper was co-authored by Haoran Li, Fan Feng, Jiarong Kang, Jin Zhang, Xiaoli Gong, Tingjuan Lu, Shuang Li, Zhe Sun, and Jordi Solé-Casals, affiliated with institutions such as Nankai University, Tsinghua University, the 903rd Hospital of the PLA, Weifang Hospital of Traditional Chinese Medicine, and Juntendo University. The paper was accepted by the journal Cognitive Computation on March 11, 2025, and published in the same year.

Research Process

1. Model Design

The core idea of the A2DM model is to guide the denoising process through artifact representation (AR). The specific workflow is as follows: - Artifact-Aware Module (AAM): First, the AAM extracts artifact representations from a pre-trained artifact classification model as prior knowledge. The AAM consists of six blocks, each containing two 1D convolutional layers and a global average pooling layer, followed by a fully connected layer to output the artifact representation. - Frequency Enhancement Module (FEM): The FEM uses a hard attention mechanism to selectively remove specific types of artifacts in the frequency domain. Specifically, the EEG signal is transformed into the frequency domain using the Fast Fourier Transform (FFT), and a binary mask is generated based on the artifact representation to select which frequency components to retain or remove. - Time-Domain Compensation Module (TCM): The TCM compensates for potential global information loss caused by the hard attention mechanism in the time domain. The TCM reconstructs the EEG signal through depth-wise convolution and 1×3 convolution, ensuring that the denoised signal retains important temporal features.

2. Datasets

The study used two datasets: - EEGDenoiseNet: This is a semi-synthetic dataset containing 4514 clean EEG segments, 3400 EOG artifact segments, and 5598 EMG artifact segments. Noisy EEG signals with multiple artifacts were generated through linear mixing. - BCI Competition IV 2A: This is a real EEG dataset containing EEG data from 9 subjects performing motor imagery tasks. The study generated training and test sets by adding noise to verify the model’s effectiveness in real-world scenarios.

3. Experiments and Evaluation

The study validated the performance of A2DM through the following steps: - Denoising Performance Evaluation: Metrics such as Root Mean Square Error (RRMSE) and Correlation Coefficient (CC) were used to evaluate the model’s denoising performance under different signal-to-noise ratios (SNRs). The results showed that A2DM significantly outperformed existing methods in removing multiple artifacts, with a 12% improvement in correlation coefficient. - Module Effectiveness Analysis: Ablation experiments were conducted to verify the roles of FEM and TCM. The results showed that removing either module led to a performance decline, indicating that these modules play complementary roles in the denoising process. - Artifact Representation Visualization: Artifact representations were visualized using t-SNE and UMAP techniques, demonstrating that the AAM effectively captures the characteristics of different types of artifacts.

Main Results

1. Denoising Performance

A2DM outperformed all comparison models on the EEGDenoiseNet dataset, especially when dealing with multiple artifacts simultaneously. Specifically, A2DM achieved RRMSE_t and RRMSE_f values of 0.6869 and 0.5314, respectively, with a correlation coefficient of 0.7248, significantly higher than other methods.

2. Module Analysis

  • Role of FEM: The FEM removes artifacts in the frequency domain using a hard attention mechanism, particularly excelling in handling EMG artifacts. Experiments showed that the FEM effectively identifies and removes artifacts distributed in the 20-30 Hz frequency range.
  • Role of TCM: The TCM compensates for potential EEG information loss caused by the hard attention mechanism in the time domain, further enhancing denoising performance.

3. Artifact Representation

By visualizing artifact representations, the study demonstrated that the AAM effectively distinguishes between different types of artifacts and provides important prior knowledge for the denoising model.

Conclusion and Significance

The A2DM model introduces artifact representation into EEG denoising tasks for the first time, successfully achieving uniform removal of multiple artifacts. The model not only significantly outperforms existing methods in denoising performance but also provides a new approach for EEG signal processing. In the future, the research team plans to explore self-supervised learning methods to generate more refined artifact representations and expand the model’s application in tasks such as event-related synchronization (ERS) and event-related desynchronization (ERD).

Research Highlights

  1. Introduction of Artifact Representation: A2DM is the first to use artifact representation as prior knowledge to guide the denoising process, significantly improving the model’s adaptability and denoising effectiveness.
  2. Application of Hard Attention Mechanism: The FEM selectively removes artifacts in the frequency domain using a hard attention mechanism, outperforming soft attention mechanisms.
  3. Complementarity of Time and Frequency Domains: The combination of TCM and FEM ensures that the denoised signal retains important information in both the time and frequency domains.
  4. Extensive Data Validation: The study validated A2DM’s effectiveness on multiple datasets, demonstrating its potential in practical applications.

Other Valuable Information

The research team also explored the performance of shallow convolutional neural networks (CNNs) in denoising tasks, finding that shallow models perform well with high-SNR signals but significantly underperform with low-SNR signals. This finding provides important insights for optimizing future EEG denoising models.

Through this study, the A2DM model not only addresses key issues in EEG signal processing but also provides new tools and methods for related research fields.