Overview and Report: Application of GRU-CNN Model Based on Microstate and Recurrence Quantification Analysis in Auditory Attention Detection

Background and Research Motivation

Attention, as a cognitive ability, plays a crucial role in the perception process, helping humans to focus on specific objects while ignoring other distractions in a complex environment. This paper investigates Auditory Attention Detection (AAD) by extracting different dynamic features from multi-channel Electroencephalogram (EEG) signals during the process of listeners focusing on a target speaker, aiming to effectively detect auditory attention in the presence of competing speakers.

Source and Author Information of the Paper

This paper was authored by Mohammadreza Eskandarinasab, Zahra Raeisi, Reza Ahmadi Lashaki, and Hamidreza Najafi, from institutions such as Utah State University, Fairleigh Dickinson University, and Tehran University of Science and Technology. The paper was published in 2024 in the journal “Scientific Reports,” with DOI: 10.1038/s41598-024-58886-y.

Research Process and Details

a) Detailed Description of the Research Process

Dataset and Preprocessing

The study utilized two publicly available databases: the DTU database and the KUL database, which contained EEG signals from 44 and 16 subjects with normal and impaired hearing, respectively. The DTU database includes ear EEG and 64-channel EEG signals sampled at 512 Hz, while the KUL database used 64-channel EEG signals sampled at 8192 Hz. Researchers preprocessed the raw EEG signals by resampling to 256 Hz, using a 0.5-70 Hz bandpass filter, and re-referencing to the average of the TP7 and TP8 electrodes.

EEG Microstate Analysis

EEG microstate analysis is used to study the temporal and spatial dynamics of brain activity. This analysis includes four stages: Global Field Power (GFP) calculation, microstate detection, microstate segmentation, and feature extraction. Each discrete EEG state is identified by the local GFP maxima of the signal and analyzed using k-means clustering to obtain optimal microstate categories. Based on this, four types of microstate features (such as occurrence frequency, duration, coverage, and average GFP) were calculated and recorded for variations among different subjects.

Recurrence Quantification Analysis (RQA)

To extract useful nonlinear dynamic features from different states of EEG signals, researchers conducted Recurrence Quantification Analysis. This method quantifies the complex and deterministic behaviors of the signals by constructing a phase space through time embedding. Extracted features include Recurrence Rate (RR) and Determinism (DET), which were used for further analysis of brain activity dynamics in EEG signals.

b) Main Results of the Study

Classification Performance of Microstate and RQA Features

The study evaluated the classification performance of microstate and RQA features in AAD using various machine learning algorithms, including k-Nearest Neighbors (KNN), Support Vector Machine (SVM), Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory Network (Bi-LSTM), and GRU-CNN based on Q-learning (GCQL). The experimental results showed that combining microstate mean GFP features with Recurrence Rate (RR) features and using the GCQL classifier significantly improved the AAD detection accuracy to 98.9%.

Feature Selection and Optimization

Through Kolmogorov-Smirnov (KS) test and Mann-Whitney U (Wilcoxon rank-sum) test, researchers found that the combined features of microstate and RQA showed significant differences in classification. Finally, by combining Recurrence Rate and mean Global Field Power, the best multivariate feature set was optimized and efficiently detected attention using the GCQL classifier.

c) Conclusion and Research Value

This study proposes a novel dynamic approach for auditory attention detection based on EEG microstate analysis and Recurrence Quantification Analysis. The research demonstrated that the GRU-CNN Q-learning model effectively identifies auditory attention from EEG data without clean auditory stimulus. Compared to existing state-of-the-art AAD methods, this approach shows superior accuracy and real-time performance. In terms of application value, this method has potential practical applications in auditory neuro-assistive devices (like neuro-guided hearing aids), capable of separating and amplifying the focused speech content for the hearing-impaired in noisy environments.

d) Research Highlights

• Innovative Methodology: First-time integration of microstate and Recurrence Quantification Analysis applied to auditory attention detection.
• High Accuracy: The proposed method achieved a detection accuracy of 98.9% within EEG segments of less than 1 second, ranking among the highest in current AAD state-of-the-art methods.
• No Clean Stimulus Requirement: Unlike forward and backward mapping methods, it does not require clean auditory stimuli, making it applicable to more real-life scenarios.

Research Significance and Future Prospects

The proposed GCQL-AAD model not only provides an efficient auditory attention detection method based on EEG signals but also showcases the potential in extracting brain activity features in complex dynamic environments. Future research can focus on further optimizing the number of electrodes to reduce computational burden while exploring applications in multi-speaker environments. The advances in auditory attention decoding from this research will offer new insights and methods for the fields of brain-computer interfaces and auditory enhancement devices.