Exploring Unsupervised Contrastive Learning Methods for ECG Analysis

Abstract

The analysis of electrocardiogram (ECG) signals is vital for the early detection and diagnosis of cardiac abnormalities, yet progress in automated ECG interpretation is constrained by the scarcity of high-quality labeled data. This thesis investigates unsupervised contrastive learning as a means to overcome these limitations and unlock clinically relevant representations from unlabeled ECG recordings. We systematically explore and compare state-of-the-art contrastive frameworks-including Momentum Contrast (MoCo) and SimCLR-adapting them for 1D ECG signals using tailored data augmentation, preprocessing, and encoder architectures. Our experimental pipeline leverages the PTB-XL dataset, employing a three-phase approach: self-supervised pre-training, supervised fine-tuning, and comprehensive evaluation using cross-validation and visualization techniques such as UMAP and K-means clustering. Results demonstrate that contrastive pre-training yields robust ECG embeddings, enabling effective downstream classification with minimal labeled data and achieving high accuracy and sensitivity in abnormality detection. Visualization of learned representations reveals the emergence of clinically meaningful clusters, validating the capacity of contrastive methods to capture latent cardiac structure. This work provides practical insights into algorithm selection, data pairing strategies, and signal formats for ECG analysis, and highlights the potential of unsupervised learning to facilitate scalable, privacy-preserving, and accessible cardiac diagnostics.