Decoding working memory-related information from repeated psychophysiological EEG experiments using convolutional and...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Extracting reliable information from electroencephalogram (EEG) signals is challenging due to low signal-to-noise ratio and intersubject variability.
• Explainable machine learning (XML) offers a novel strategy to overcome these limitations in analyzing complex brain data.

Purpose of the Study:

• To evaluate the efficacy of XML for classifying and decoding electrical brain activity related to information retention.
• To identify reliable neural features associated with working memory using diverse deep learning models.

Main Methods:

• Designed four distinct neural network models with varying architectures, training strategies, and input representations.
• Classified single experimental trials from a working memory task using these models.
• Employed input perturbation analysis to determine the importance of EEG channels and frequency bands.

Main Results:

• The best models achieved an accuracy of 65.29% and a Matthews correlation coefficient of 0.288, outperforming a reference model.
• A significant correlation (0.36, p=0.0007) was found between classification scores and behavioral performance.
• A common subset of features identified by all models aligned with known neurophysiological processes critical for attention and working memory.

Conclusions:

• Explainable deep learning is a potent tool for decoding information from EEG signals.
• Analyzing a range of models is essential for identifying stable and reliable neural features.
• The study underscores the necessity of explainable modeling, as high-accuracy models may inadvertently rely on artifactual data.