EEGminer: discovering interpretable features of brain activity with learnable filters

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Brain activity patterns can indicate brain states and predict behavior.
• Identifying relevant features from electroencephalography (EEG) data is challenging.
• Need for systems that learn informative representations from multichannel EEG recordings.

Purpose of the Study:

• To design a system for learning interpretable latent representations from ongoing EEG activity.
• To develop a differentiable decoding pipeline with learnable filters and feature extraction.
• To enable stable end-to-end model training and uncover meaningful EEG features.

Main Methods:

• Proposed a novel differentiable decoding pipeline.
• Utilized learnable filters parameterized by generalized Gaussian functions for smooth derivatives.
• Employed signal magnitude and functional connectivity for feature extraction.

Main Results:

• Demonstrated utility on a large EEG dataset (721 subjects) for music perception analysis.
• Identified consistent trends and individual differences in music perception.
• Achieved comparable accuracy to black-box models while offering strong feature interpretability.
• Showcased applications in emotion recognition (SEED) and workload classification.

Conclusions:

• The proposed method provides interpretable features from EEG data.
• Achieves high accuracy, comparable to deep learning models.
• Enhances trustworthiness of deep learning models for practical neuroscience applications.
• Offers new insights into functional connectivity during music perception.