Deep Correlation Analysis for Audio-EEG Decoding.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Non-invasive electroencephalography (EEG) is crucial for recording brain activity but is susceptible to artifacts.
• Existing methods for stimulus-response analysis in EEG often rely on linear techniques, which can be limited.
• Artifacts in EEG recordings can negatively impact the accuracy of stimulus-response correlation analysis.

Purpose of the Study:

• To develop a novel neural network-based framework for correlation analysis of electroencephalography (EEG) data.
• To improve the accuracy of stimulus-response correlation analysis in the presence of EEG artifacts.
• To enhance both intra-subject and inter-subject stimulus-response correlations using deep learning models.

Main Methods:

• Proposed a deep neural network model for intra-subject audio-EEG analysis, optimizing correlation loss directly.
• Developed a neural network model with a shared encoder architecture for inter-subject stimulus-response correlation analysis.
• The models were designed to suppress EEG artifacts while preserving stimulus-related neural components.

Main Results:

• The proposed deep learning models significantly improved Pearson correlation compared to traditional linear methods.
• Average absolute improvements in correlation were 7.4% for speech stimuli and 29.3% for music stimuli.
• Analysis demonstrated the effectiveness of the models in artifact suppression and preservation of stimulus-related information.

Conclusions:

• Neural network-based correlation analysis offers a significant advancement over linear methods for EEG stimulus-response analysis.
• The proposed deep models effectively mitigate EEG artifacts, leading to more accurate brain-stimulus correlation.
• This framework holds promise for more reliable non-invasive brain activity analysis, particularly for auditory stimuli.