Attention-Guided Graph Structure Learning Network for EEG-enabled Auditory Attention Detection

Neurology Computer Science Engineering Electrical Engineering Information Science Electronic Information Artificial Intelligence Medicine
auditory attention detection electroencephalography graph structure learning graph convolutional network neuro-steered hearing aids

Application of Attention-guided Graph Structure Learning Network for EEG-enabled Auditory Attention Detection

Attention-guided Graph Structure Learning Network

Academic Background

The “cocktail party effect” describes the human brain’s ability to selectively concentrate attention on one speaker while ignoring others in a multi-talker environment. However, for individuals with hearing impairments, this situation poses a significant challenge. Although modern hearing aids, such as hearing aids and cochlear implants, are effective in noise reduction, they often cannot distinguish the signal that the listener needs to focus on. The auditory attention detection (AAD) task has the potential to solve this problem by directly extracting attention-related information from the brain. Neuroscience research has shown that non-invasive neural recording techniques, such as electroencephalography (EEG), have great potential for decoding auditory attention. To address the issue of EEG signal decoding, researchers have developed various methods to interpret EEG signals and thereby determine attention, adjusting the performance of hearing aids.

Paper Source and Author Information

This paper titled “Attention-guided graph structure learning network for EEG-enabled auditory attention detection” is written by Xianzhang Zeng, Siqi Cai, and Longhan Xie, affiliated with the Department of Electrical and Computer Engineering at Southern University of Science and Technology in Guangzhou, China, and the National University of Singapore, respectively. The paper, published in the Journal of Neural Engineering in 2024, provides a detailed decoding of how to use EEG signals for auditory attention detection.

Detailed Research Process

Workflow

The study proposes a novel attention-guided graph structure learning network (AGSLEnet) that leverages the inherent relationships between EEG signals to improve AAD performance. For this network, the research was conducted through several steps with thorough analysis and experimentation.

  1. Multi-channel EEG Recording and Preprocessing: EEG signals are first referenced to the average response of all channels. Then, bandpass filtering is performed from 1-32Hz, and the filtered EEG signals are downsampled to 128Hz. Additionally, independent component analysis (ICA) is performed using the EEGLAB toolbox to reduce artifact effects. After these processes, a series of EEG slices called Decision Windows are obtained.

  2. Temporal Feature Extraction: In the temporal feature extraction module, one-dimensional convolutional layers and Exponential Linear Unit (ELU) activation functions are employed, combined with Batch Normalization (BN) layers to aggregate temporal information from each EEG channel.

  3. Attention-guided Graph Representation: The study constructs an attention-based graph representation. Specifically, linear projections are used to transform feature maps into query vectors and key vectors, and their dot products are calculated to obtain an attention weight matrix, which is then used to dynamically generate the adjacency matrix of the EEG signals, capturing durable associations.

  4. Graph Convolution: Graph convolution operations are computed using spectral filters derived from the Normalized Laplacian Matrix, capturing global information from the EEG graph. Graph convolution enables the network to extend convolution operations to the graph domain, achieving signal processing through Fourier transforms.

  5. End-to-end AAD Classifier: Finally, AGSLEnet is an end-to-end system that takes multi-channel EEG signals as input and outputs a binary attention judgment. In this step, the feature maps processed by Temporal Average Pooling and Flattening are passed through a Fully Connected Layer with a softmax activation function for the final binary classification.

Subjects and Experiments

To evaluate the effectiveness of AGSLEnet, extensive experiments were conducted on two publicly available AAD datasets, the KUL dataset and the DTU dataset. In each dataset, participants’ multi-channel EEG signals were recorded while they listened to and focused on specific speakers’ voices.

  1. KUL Dataset: Includes EEG data from 16 normal-hearing individuals who were instructed to selectively attend to one speaker in a dual-talker scenario. 64-channel EEG signals were recorded in an acoustically and electromagnetically shielded room at a sampling rate of 8192 Hz.

  2. DTU Dataset: Collected from 18 normal-hearing participants, recording their EEG signals in simulated reverberant and anechoic environments. EEG signals were recorded using a BioSemi active system at a sampling rate of 512 Hz.

Main Research Results

  1. Effectiveness of the Attention-guided Graph Structure Learning Network (AGSLEnet): AGSLEnet demonstrated superior AAD performance on both the KUL and DTU datasets. By constructing an attention-based dynamic graph representation, AGSLEnet successfully captured the inherent relationships between EEG signals, resulting in significantly higher AAD accuracy compared to other competing models.

  2. AAD Performance in Low-latency Scenarios: Within the decision window length range of 0.1 seconds to 2 seconds, AGSLEnet exhibited outstanding AAD accuracy. For example, at a low-latency scenario of 0.1 seconds, the accuracy reached 88.1%, and for a 1-second decision window, the accuracy was 93.6%.

  3. Comparative Study: By comparing AGSLEnet with other models (such as CNN, RGC, etc.), the results showed that AGSLEnet outperformed the others across all decision window lengths. For instance, compared to other models, AGSLEnet increased accuracy by 3.5% to 9.5% in the 1-second decision window.

Conclusion and Significance

This research, by proposing the attention-guided graph structure learning framework AGSLEnet, provides new scientific insights and practical possibilities for auditory attention detection using EEG signals. Not only does AGSLEnet outperform traditional methods in terms of AAD accuracy, but it also demonstrates the effectiveness of dynamically constructing graph structures from EEG signals. This novel neural decoding technique has the potential to drive the development of neural-guided audible devices and provide new tools and methods for real-world applications.

Research Highlights

  1. Innovative Method: Utilizes the attention mechanism to dynamically generate graph structures from EEG signals, optimizing attention decoding performance.
  2. Extensive Experiments: Validates the effectiveness and generalization ability of the model through extensive experiments on two public datasets, KUL and DTU.
  3. Low-latency Applications: Performs well in various low-latency scenarios, laying the foundation for real-time neural-guided device applications.
  4. Interdisciplinary Significance: Provides new research insights, helping to deepen the understanding of brain functional connectivity and auditory attention mechanisms.

Additional Information Worth Noting

Future research can further explore the application of AGSLEnet on more realistic datasets, particularly those including data samples from multiple environments. By expanding the research scope, not only can the theoretical research results be validated, but also the model’s performance in specific applications can be enhanced. Additionally, the application of self-supervised learning (SSL) techniques in EEG analysis is also worth anticipating.

Through the implementation of the AGSLEnet framework, this research has broad academic and application prospects in auditory attention detection, low-latency scenario applications, EEG signal decoding, and brain function research.