[Paper Review] CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients

Kiyasseh, Dani, Zhu, Tingting+1 more|arXiv (Cornell University)|May 27, 2020

Phonocardiography and Auscultation Techniques38 citations

TL;DR

CLOCS introduces patient-centric contrastive learning for ECGs, leveraging temporal, spatial, and patient-level invariances to learn robust representations that outperform BYOL and SimCLR, with strong data efficiency.

ABSTRACT

The healthcare industry generates troves of unlabelled physiological data. This data can be exploited via contrastive learning, a self-supervised pre-training method that encourages representations of instances to be similar to one another. We propose a family of contrastive learning methods, CLOCS, that encourages representations across space, time, \ extit{and} patients to be similar to one another. We show that CLOCS consistently outperforms the state-of-the-art methods, BYOL and SimCLR, when performing a linear evaluation of, and fine-tuning on, downstream tasks. We also show that CLOCS achieves strong generalization performance with only 25\\% of labelled training data. Furthermore, our training procedure naturally generates patient-specific representations that can be used to quantify patient-similarity.

Motivation & Objective

• Motivate learning from large unlabelled ECG data where labels are scarce in clinical settings.
• Develop patient-specific contrastive learning methods that exploit temporal and spatial invariances in ECG signals.
• Show that CLOCS methods outperform state-of-the-art self-supervised baselines on arrhythmia classification tasks.
• Demonstrate data-efficient training with limited labelled data and emergent patient-specific representations.
• Explore the generalization and transferability of learned representations across datasets and downstream tasks.

Proposed method

• Define positive pairs not only from same instance but also from same patient across transformations.
• Introduce three transformation-based coding schemes: Contrastive Multi-segment Coding (CMSC), Contrastive Multi-lead Coding (CMLC), and Contrastive Multi-segment Multi-lead Coding (CMSMLC).
• Use patient-specific Noise Contrastive Estimation (NCE) loss with cosine similarity and temperature scaling to maximize agreement within the same patient and minimize it across patients.
• Leverage temporal invariance (segment-level) and spatial invariance (lead-level) to create augmented views of ECG signals.
• Evaluate under linear evaluation and fine-tuning protocols on multiple ECG datasets with varying lead counts (4 leading configurations and 12-lead setups).
• Compare CLOCS against Random Init., Supervised, MT-SSL, BYOL, and SimCLR baselines.

Experimental results

Research questions

• RQ1 Can patient-level context improve contrastive representations for ECG signals beyond instance-level context?
• RQ2 Do temporal and spatial invariances in ECGs jointly improve representation learning when combined with patient-specific positive pairs?
• RQ3 How do CMSC, CMLC, and CMSMLC compare to standard self-supervised baselines on arrhythmia classification tasks?
• RQ4 Is CLOCS data-efficient, enabling strong downstream performance with limited labelled data?
• RQ5 Do learned representations generalize across datasets and enable patient-centric analyses?

Key findings

• CLOCS variants consistently outperform BYOL and SimCLR in linear evaluation and fine-tuning on arrhythmia classification tasks.
• CMSC achieves strong linear evaluation performance (example: Chapman dataset) and outperforms several baselines across datasets.
• Using 25% labelled data, CLOCS representations remain competitive, indicating strong data efficiency.
• CMSC and CMSMLC generate more separable, patient-specific representations compared to SimCLR, evidenced by intra- vs inter-patient distance analyses.
• Across perturbation studies, CMSC maintains superior performance over SimCLR, suggesting benefits arise from patient-level shared context rather than perturbation choices alone.
• All CLOCS variants outperform SimCLR in the majority of experiments, with CMSC leading when using 12 leads.