[Paper Review] Uncovering the structure of clinical EEG signals with self-supervised learning

Hubert Banville, Omar Chehab+3 more|arXiv (Cornell University)|Jul 31, 2020

EEG and Brain-Computer Interfaces95 references231 citations

TL;DR

The paper investigates self-supervised learning (SSL) for EEG, introducing three pretext tasks (relative positioning, temporal shuffling, and contrastive predictive coding) to learn EEG representations and evaluates them on sleep staging and pathology detection, showing SSL features outperform supervised baselines in low-label regimes and reveal physiologically meaningful structure.

ABSTRACT

Objective. Supervised learning paradigms are often limited by the amount of labeled data that is available. This phenomenon is particularly problematic in clinically-relevant data, such as electroencephalography (EEG), where labeling can be costly in terms of specialized expertise and human processing time. Consequently, deep learning architectures designed to learn on EEG data have yielded relatively shallow models and performances at best similar to those of traditional feature-based approaches. However, in most situations, unlabeled data is available in abundance. By extracting information from this unlabeled data, it might be possible to reach competitive performance with deep neural networks despite limited access to labels. Approach. We investigated self-supervised learning (SSL), a promising technique for discovering structure in unlabeled data, to learn representations of EEG signals. Specifically, we explored two tasks based on temporal context prediction as well as contrastive predictive coding on two clinically-relevant problems: EEG-based sleep staging and pathology detection. We conducted experiments on two large public datasets with thousands of recordings and performed baseline comparisons with purely supervised and hand-engineered approaches. Main results. Linear classifiers trained on SSL-learned features consistently outperformed purely supervised deep neural networks in low-labeled data regimes while reaching competitive performance when all labels were available. Additionally, the embeddings learned with each method revealed clear latent structures related to physiological and clinical phenomena, such as age effects. Significance. We demonstrate the benefit of self-supervised learning approaches on EEG data. Our results suggest that SSL may pave the way to a wider use of deep learning models on EEG data.

Motivation & Objective

• Motivate SSL as a solution to the scarcity of labeled EEG data in clinical settings.
• Introduce three EEG-specific SSL pretext tasks to learn robust representations from unlabeled data.
• Benchmark SSL representations on sleep staging and pathology detection against supervised and hand-crafted baselines.
• Analyze the learned embeddings for physiologically and clinically meaningful structure.

Proposed method

• Define three SSL pretext tasks for EEG: relative positioning (RP), temporal shuffling (TS), and contrastive predictive coding (CPC).
• Use end-to-end trainable encoders h_Θ to map EEG windows to feature space, with contrastive modules g_RP, g_TS, or parametric CPC components.
• Optimize using binary logistic loss for RP/TS and InfoNCE loss for CPC, training end-to-end with a downstream linear classifier or logistic regression.
• Evaluate two EEG architectures (StagerNet and ShallowNet) as embedders, with a GRU-based autoregressive component for CPC.
• Compare SSL with baselines: random weights, convolutional autoencoders, purely supervised models, and handcrafted features.
• Conduct experiments on sleep staging (Physionet Challenge 2018) and pathology detection (TUH Abnormal EEG).

Experimental results

Research questions

• RQ1 What SSL tasks best capture relevant structure in EEG data?
• RQ2 How do SSL features compare to unsupervised and supervised baselines in downstream EEG classification tasks?
• RQ3 What characteristics do SSL-learned embeddings reveal about physiological and clinical phenomena (e.g., age effects)?

Key findings

• SSL features enable linear classifiers to outperform purely supervised networks in low-labeled data regimes.
• SSL representations reach competitive performance with full label supervision when all labels are available.
• Embeddings from SSL methods reveal latent structure related to physiological and clinical factors such as age.
• Two clinically relevant EEG tasks—sleep staging and pathology detection—benefit from SSL over baseline approaches.
• The study provides evidence that SSL can enhance deep learning applicability to EEG data in clinical contexts.