Adversarial Spatiotemporal Contrastive Learning for Electrocardiogram Signals.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Machine Learning

Background:

• Extracting invariant representations from unlabeled electrocardiogram (ECG) signals using deep neural networks (DNNs) presents a significant challenge.
• While contrastive learning offers a promising approach for unsupervised learning, it requires enhancement in robustness to noise and the ability to learn spatiotemporal and semantic representations akin to cardiologists.

Purpose of the Study:

• To propose a novel patient-level adversarial spatiotemporal contrastive learning (ASTCL) framework for unsupervised ECG representation learning.
• To improve the robustness of DNNs to noise and enable the learning of comprehensive spatiotemporal and semantic ECG representations.

Main Methods:

• The ASTCL framework incorporates ECG augmentations (noise enhancement and denoising) to bolster noise robustness.
• An adversarial module, employing a game between a discriminator and encoder, learns invariant representations by discarding perturbations.
• A spatiotemporal contrastive module combines spatiotemporal prediction and patient discrimination, utilizing patient-level positive pairs to learn category representations and avoid model collapse.

Main Results:

• The proposed ASTCL framework demonstrated enhanced robustness to noise in ECG signals.
• The method effectively learned spatiotemporal and semantic representations from unlabeled ECG data.
• Experiments on multiple ECG datasets showed that ASTCL outperforms current state-of-the-art methods.

Conclusions:

• The ASTCL framework provides an effective solution for unsupervised learning of invariant ECG representations.
• The proposed method advances the capability of DNNs in analyzing complex ECG data, offering potential for improved diagnostic tools.