A Framework for Systematically Evaluating the Representations Learned by A Deep Learning Classifier from Raw Multi-Channel Electroencephalogram Data

Charles A. Ellis, Abhinav Sattiraju, Robyn L. Miller, Vince D. CalhounbioRxiv 2023 · DOI

TL;DR

This paper introduces a novel framework and taxonomy for explaining deep learning models trained on raw EEG data, specifically addressing the challenge of "black box" models. Their framework provides unprecedented insights into spatio-spectral interactions within EEG data, a crucial step for understanding neurological and psychiatric disorders.

What Problem Does This Paper Address?

Deep learning models are increasingly used for analyzing raw electroencephalogram (EEG) data, offering potential performance gains over traditional methods that rely on manually engineered features. However, this shift to raw data often results in a loss of interpretability, making it difficult to understand why a model makes a particular prediction. While existing explainability methods exist for EEG, they often focus on individual features (spatial, spectral, or temporal) and lack the ability to uncover complex interactions between these features, which are critical for understanding brain function and dysfunction. This paper aims to bridge this gap by providing a systematic framework for evaluating deep learning models on raw EEG, with a particular focus on uncovering spatio-spectral interactions.

Methods

The authors propose a comprehensive framework for EEG explainability, building upon existing techniques and introducing novel approaches. Key methods include:

  • Taxonomy of Explainability Methods: A structured categorization of existing methods for raw EEG explainability, highlighting gaps in current capabilities.
  • Layer-wise Relevance Propagation (LRP): Used extensively to determine the importance of individual channels (spatial importance) and to understand how features contribute to model predictions.
  • Ablation Studies: Used in conjunction with LRP to investigate spatial interactions, by removing data from specific channels and observing the impact on relevance. This helps identify how the model uses information from one channel to interpret another.
  • Spectral Perturbation: Combined with LRP to analyze spatio-spectral importance, by zeroing out specific frequency bands within channels and observing the change in relevance.
  • Spatio-Spectral Interaction Analysis: A novel approach that extends spectral perturbation to examine how perturbing frequency bands in one channel affects the relevance of other channels.
  • Prototyping and Clustering: Used for temporal importance, this involves dimensionality reduction (PCA) and clustering of model activations to identify representative samples and their corresponding important waveforms.

Datasets

The study utilizes a publicly available scalp EEG dataset.

  • Dataset: A publicly available scalp EEG dataset consisting of resting-state EEG data from 28 healthy controls (HCs) and 30 individuals with Major Depressive Disorder (MDD). This dataset has been used in previous studies.

Key Insights and Ideas

  • The Importance of Interactions: The paper strongly emphasizes that understanding interactions between different EEG features (spatial, spectral) is crucial for a deeper understanding of brain function and disorders, and that existing methods often overlook these.
  • Spatio-Spectral Interactions are Novel: The authors claim their spatio-spectral interaction analysis is a first for raw EEG deep learning explainability, offering a new lens into how models process complex brain signals.
  • MDD-Specific Patterns Emerge: The framework successfully identified distinct patterns of electrode and frequency band importance and interactions between healthy individuals and those with MDD.
  • Frontal and Central Electrodes are Key for MDD: The model relied heavily on frontal and central electrodes, particularly with specific frequency bands (α, β, γlow), for identifying MDD.
  • Posterior Electrodes are Important for HCs: For healthy controls, the model showed a greater reliance on posterior electrodes, especially with δ and θ frequencies.

Key Results

  • High Model Performance: The trained 1D-CNN achieved high classification performance, with mean ACC and BACC above 80%. The model used for explainability achieved 100% performance on all metrics.
  • Spatial Importance: Certain electrodes were consistently more important than others. For example, F4, T5, and O2 were generally important for both classes, while F3, F4, Cz, C4, T5, T6, and O2 showed above-average importance.
  • Spatial Interactions: Significant interactions were found, with HCs showing inter-occipital interactions (O1 and O2) and MDDs showing interactions involving Cz and Fp2. The study noted that MDDs tended to have more non-significant negative interactions than HCs.
  • Spatio-Spectral Importance: The model relied more on θ and α for HCs, and β and γlow for MDDs, across a wider range of channels. Frontal electrodes were important for both classes but for different reasons (e.g., frontal δ and θ for HCs, frontal β and γlow for MDDs).
  • Spatio-Spectral Interactions: Interactions were widespread, with HCs showing more posterior interactions and MDDs showing more anterior interactions.
  • Temporal Importance: The analysis identified representative samples and highlighted specific waveforms and temporal patterns that differentiated MDDs from HCs. For instance, HCs showed consistent δ and θ oscillations, while MDDs showed more β oscillations in specific channels.

Limitations and Future Work

  • Dataset Size: The authors acknowledge that a larger dataset would be needed to make generalizable claims about feature importance for MDD diagnosis.
  • Single Perturbation: For spatial and spatio-spectral interactions, features were perturbed only once. Future work could explore multiple perturbations to better understand their effects.
  • Computational Complexity: The computational cost of repeatedly applying LRP and perturbation methods limits their application to very large datasets or complex models. Future work could explore more computationally efficient methods or adapt techniques like SHAP.
  • Line Noise: The study used data where line noise was filtered, preventing the use of line noise-related ablation for spatial interactions. Future studies could explore this if raw data with line noise is available.
  • Architecture Specificity: Some existing explainability methods require specially designed architectures. The authors aim for their temporal importance approach to be applicable to a wider range of deep learning classifiers.
  • Future Directions: The authors suggest expanding their prototyping approach, applying more local explainability methods, adapting techniques like SHAP, exploring model attention mechanisms, and using inverse modeling to obtain source space signals for enhanced insights.

Summary updated

Abstract

ABSTRACT The application of deep learning methods to raw electroencephalogram (EEG) data is growing increasingly common. While these methods offer the possibility of improved performance relative to other approaches applied to manually engineered features, they also present the problem of reduced explainability. As such, a number of studies have sought to provide explainability methods uniquely adapted to the domain of deep learning-based raw EEG classification. In this study, we present a taxonomy of those methods, identifying existing approaches that provide insight into spatial, spectral, and temporal features. We then present a novel framework consisting of a series of explainability approaches for insight into classifiers trained on raw EEG data. Our framework provides spatial, spectral, and temporal explanations similar to existing approaches. However, it also, to the best of our knowledge, proposes the first explainability approaches for insight into spatial and spatio-spectral interactions in EEG. This is particularly important given the frequent use and well-characterized importance of EEG connectivity measures for neurological and neuropsychiatric disorder analysis. We demonstrate our proposed framework within the context of automated major depressive disorder (MDD) diagnosis, training a high performing one-dimensional convolutional neural network with a robust cross-validation approach on a publicly available dataset. We identify interactions between frontal and central electrodes and other electrodes and identify differences in frontal δ, θ, β, and γ low between healthy controls and individuals with MDD. Our study represents a significant step forward for the field of deep learning-based raw EEG classification, providing new capabilities in interaction explainability and providing direction for future innovations through our proposed taxonomy.
BibTeX
@misc{ellis2023framework,
  author = {Charles A. Ellis and Abhinav Sattiraju and Robyn L. Miller and Vince D. Calhoun},
  title = {{A Framework for Systematically Evaluating the Representations Learned by A Deep Learning Classifier from Raw Multi-Channel Electroencephalogram Data}},
  year = {2023},
  doi = {10.1101/2023.03.20.533467},
  howpublished = {bioRxiv},
}
Figure