[Paper Review] Representation learning for improved interpretability and classification accuracy of clinical factors from EEG

Garrett Honke, Irina Higgins+7 more|arXiv (Cornell University)|Oct 28, 2020

EEG and Brain-Computer Interfaces52 references4 citations

TL;DR

This paper proposes a deep representation learning framework that uses disentangled β-VAE to denoise EEG signals and a Symbol–Concept Association Network (SCAN) to interpret clinical factors from EEG data. It improves classification accuracy over hand-engineered baselines and enables interpretable, reusable representations for age, gender, depression, and other clinical traits, while automatically recovering known biomarkers like anhedonia-related ERP patterns.

ABSTRACT

Despite extensive standardization, diagnostic interviews for mental health disorders encompass substantial subjective judgment. Previous studies have demonstrated that EEG-based neural measures can function as reliable objective correlates of depression, or even predictors of depression and its course. However, their clinical utility has not been fully realized because of 1) the lack of automated ways to deal with the inherent noise associated with EEG data at scale, and 2) the lack of knowledge of which aspects of the EEG signal may be markers of a clinical disorder. Here we adapt an unsupervised pipeline from the recent deep representation learning literature to address these problems by 1) learning a disentangled representation using $β$-VAE to denoise the signal, and 2) extracting interpretable features associated with a sparse set of clinical labels using a Symbol-Concept Association Network (SCAN). We demonstrate that our method is able to outperform the canonical hand-engineered baseline classification method on a number of factors, including participant age and depression diagnosis. Furthermore, our method recovers a representation that can be used to automatically extract denoised Event Related Potentials (ERPs) from novel, single EEG trajectories, and supports fast supervised re-mapping to various clinical labels, allowing clinicians to re-use a single EEG representation regardless of updates to the standardized diagnostic system. Finally, single factors of the learned disentangled representations often correspond to meaningful markers of clinical factors, as automatically detected by SCAN, allowing for human interpretability and post-hoc expert analysis of the recommendations made by the model.

Motivation & Objective

• To address the lack of automated, scalable methods for denoising noisy EEG data in clinical settings.
• To identify interpretable, biologically meaningful EEG features linked to clinical factors such as depression and age.
• To develop a reusable representation that supports fast re-mapping to new clinical labels without retraining.
• To improve classification accuracy of clinical factors using EEG data beyond traditional hand-engineered baselines.
• To enable post-hoc interpretability of model decisions through disentangled latent factors associated with specific clinical traits.

Proposed method

• A β-VAE is trained on averaged EEG trajectories (Event-Related Potentials, ERPs) to learn a disentangled, low-dimensional representation that denoises raw EEG signals.
• The disentangled latent space from β-VAE is used to train a Symbol–Concept Association Network (SCAN), which learns to map clinical labels (e.g., depression, gender) to specific latent dimensions.
• SCAN uses a conditional encoder to infer latent codes for clinical labels and a decoder to reconstruct ERPs from these codes, enabling visualization of EEG patterns associated with each label.
• The model enables zero-shot or few-shot re-mapping to new clinical labels by reusing the pre-trained β-VAE and fine-tuning only the SCAN head.
• Latent factors are evaluated for interpretability by analyzing which dimensions correspond to specific labels and visualizing their traversals and posterior distributions.
• Classification performance is evaluated using L2-regularized logistic regression on latent codes, with sparsity and precision-recall metrics used to assess interpretability and accuracy.

Experimental results

Research questions

• RQ1 Can disentangled representation learning from EEG data improve classification accuracy of clinical factors such as depression and age compared to canonical hand-engineered baselines?
• RQ2 Can the learned disentangled representations be used to automatically extract denoised ERPs from single-trial EEG trajectories?
• RQ3 Do individual latent dimensions in the β-VAE representation correspond to interpretable, clinically meaningful features such as gender or depression status?
• RQ4 Can the model support fast, supervised re-mapping to new clinical labels without retraining the entire representation pipeline?
• RQ5 Can the model recover known neurophysiological biomarkers, such as those related to anhedonia, from EEG data?

Key findings

• The proposed method outperformed the canonical hand-engineered baseline (LPP representation) in balanced classification accuracy for age, gender, depression, and Axis 1 diagnosis.
• SCAN achieved a sparsity of 87.5% in feature usage, significantly higher than the 3.25% sparsity of L1-regularized logistic regression, indicating more interpretable and focused feature selection.
• The model successfully reconstructed ERP patterns corresponding to depression status: participants with depression showed no difference in ERP response to positive vs. neutral stimuli, while non-depressed participants showed a clear distinction.
• Disentangled latent dimensions were found to correspond to specific clinical labels—e.g., gender was encoded in a single latent dimension (z₅), with distinct, non-overlapping Gaussian distributions for males and females.
• The UDR score (a proxy for disentanglement quality) showed a significant positive correlation (r = 0.33, p < .001) with classification accuracy, indicating that better disentanglement leads to better performance.
• The framework enabled automatic denoising of single-trial EEG trajectories via the β-VAE decoder, producing clean ERP-like waveforms without requiring averaging across trials.