[Paper Review] Subject-Aware Contrastive Learning for Biosignals

Joseph Y. Cheng, Hanlin Goh+3 more|arXiv (Cornell University)|Jun 30, 2020

Muscle activation and electromyography studies45 references77 citations

TL;DR

A self-supervised contrastive learning approach for biosignals that incorporates subject-aware losses to handle inter-subject variability, with extensive EEG and ECG evaluations showing competitive performance and improved representations when labeled data and subjects are limited.

ABSTRACT

Datasets for biosignals, such as electroencephalogram (EEG) and electrocardiogram (ECG), often have noisy labels and have limited number of subjects (<100). To handle these challenges, we propose a self-supervised approach based on contrastive learning to model biosignals with a reduced reliance on labeled data and with fewer subjects. In this regime of limited labels and subjects, intersubject variability negatively impacts model performance. Thus, we introduce subject-aware learning through (1) a subject-specific contrastive loss, and (2) an adversarial training to promote subject-invariance during the self-supervised learning. We also develop a number of time-series data augmentation techniques to be used with the contrastive loss for biosignals. Our method is evaluated on publicly available datasets of two different biosignals with different tasks: EEG decoding and ECG anomaly detection. The embeddings learned using self-supervision yield competitive classification results compared to entirely supervised methods. We show that subject-invariance improves representation quality for these tasks, and observe that subject-specific loss increases performance when fine-tuning with supervised labels.

Motivation & Objective

• Motivate learning robust biosignal representations with limited labels and few subjects.
• Address intersubject variability by introducing subject-aware mechanisms in SSL.
• Develop domain-inspired data augmentations tailored to biosignals.
• Integrate subject information via (a) subject-specific contrastive loss and (b) adversarial subject invariance.
• Evaluate representations on EEG decoding and ECG anomaly detection tasks.

Proposed method

• Use an encoder G to map augmented biosignal segments to latent representations.
• Optimize mutual information between two augmentations via InfoNCE contrastive loss.
• Introduce subject-aware training: (i) subject-specific contrastive loss, and (ii) adversarial training to promote subject invariance.
• Incorporate subject information into negative sampling and/or regularization with a controllable weight lambda.
• Develop and apply time-series augmentations (temporal cutout, temporal delays, noise, bandstop, signal mixing) and EEG-specific spatial augmentations.
• Train with momentum encoders to enlarge the set of negatives across batches and time.

Experimental results

Research questions

• RQ1 Can self-supervised contrastive learning produce discriminative biosignal representations with limited labels and subjects?
• RQ2 Does incorporating subject-awareness (invariance or specificity) improve downstream EEG/ECG tasks compared to base SSL?
• RQ3 Which data augmentations best preserve task-relevant information for biosignals in a contrastive framework?
• RQ4 How do subject-aware SSL representations fare when fine-tuned with supervised labels on unseen subjects?

Key findings

• Subject-invariant SSL improves representation quality for EEG and ECG when subjects are limited.
• Subject-specific SSL and subject-invariant SSL can both reduce subject-identification information in embeddings and improve downstream accuracy.
• Temporal augmentations (cutout, delays) most enhance EEG representation learning among tested transformations.
• Subject-aware SSL yields competitive EEG decoding and ECG beat/rhythm classification compared to fully supervised baselines, especially with few labels or subjects.
• Fine-tuning from SSL initializations (especially subject-specific) improves end-to-end supervised performance on EEG tasks.
• For ECG, subject-invariant SSL is beneficial in low-label/low-subject regimes and can require tuning of the regularization parameter lambda to balance subject information.