[Paper Review] CLOCS: Contrastive Learning of Cardiac Signals.

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

Phonocardiography and Auscultation Techniques13 references12 citations

TL;DR

CLOCS proposes a contrastive learning framework for cardiac signals that learns representations by aligning embeddings across time, leads, and patients, significantly outperforming SimCLR on downstream tasks and achieving strong performance with only 25% of labeled data, while enabling patient-specific representation learning for patient similarity quantification.

ABSTRACT

The healthcare industry generates troves of unlabelled physiological data. This data can be exploited via contrastive learning, a self-supervised pre-training mechanism that encourages representations of instances to be similar to one another. We propose a family of contrastive learning methods, CLOCS, that encourages representations across time, leads, and patients to be similar to one another. We show that CLOCS consistently outperforms the state-of-the-art approach, SimCLR, on both linear evaluation and fine-tuning 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

• To address the challenge of limited labeled cardiac signal data by leveraging self-supervised contrastive learning.
• To improve representation learning in ECG data by aligning embeddings across temporal segments, leads, and patients.
• To achieve strong downstream performance with minimal labeled data, particularly 25% of the training set.
• To generate patient-specific representations that enable quantification of patient similarity.

Proposed method

• CLOCS employs a contrastive learning objective that encourages similarity between positive pairs of views derived from the same ECG signal.
• Views are generated by augmenting the same ECG signal across time (temporal crops), leads (channel-level augmentations), and patients (cross-patient views).
• The model learns a shared representation space where positive pairs (e.g., different augmentations of the same signal) are pulled closer, while negative pairs are pushed apart.
• The contrastive loss is optimized using a momentum encoder and a queue-based memory bank to stabilize training.
• The framework supports both linear evaluation and fine-tuning on downstream tasks such as arrhythmia detection.
• Patient-specific representations are extracted from the final layer and used to compute patient similarity via cosine similarity.

Experimental results

Research questions

• RQ1 Can contrastive learning across time, leads, and patients improve representation quality in ECG signals compared to state-of-the-art methods?
• RQ2 How well does CLOCS generalize when only 25% of the labeled data is used for downstream fine-tuning?
• RQ3 Can the learned representations effectively quantify patient similarity in a clinically meaningful way?
• RQ4 Does CLOCS outperform SimCLR in both linear evaluation and fine-tuning settings on ECG classification tasks?

Key findings

• CLOCS consistently outperforms SimCLR on both linear evaluation and fine-tuning downstream tasks across multiple ECG datasets.
• CLOCS achieves strong performance with only 25% of the labeled training data, demonstrating robust generalization under data scarcity.
• The model generates patient-specific representations that enable meaningful patient-similarity quantification through cosine similarity of learned embeddings.
• The contrastive learning objective across time, leads, and patients leads to more robust and generalizable representations than single-modality or single-patient approaches.