Reading Your Heart: Learning ECG Words and Sentences via Pre-training ECG Language Model

Electrocardiograms (ECGs) are essential for the clinical diagnosis of arrhythmias and other heart diseases, but deep learning methods based on ECGs often face limitations due to the need for highquality annotations. Although previous ECG self-supervised learning (eSSL) methods have made significant progress, they typically treat ECG signals as general time-series data, using fixed steps and window sizes, which often ignore the heartbeat and rhythmic characteristics and potential semantic relationships in ECG signals. In this work, we introduce a novel perspective on ECG signals, treating heartbeats as words and rhythms as sentences. Based on this perspective, we propose HeartLang, a novel self-supervised learning framework for ECG language processing. Within this framework, we construct an ECG vocabulary and pre-train the model using masked prediction on ECG sentences to learn both heartbeat-level and rhythm-level representations, uncovering the latent semantic relationships in ECG signals. We also developed three parameter scales for HeartLang, namely, HeartLang-Small, HeartLang-Base, and HeartLang-Large, and conducted pre-training and downstream task testing on the standard benchmark dataset PTB-XL.The experimental results demonstrate that our method exhibits superior performance compared to other eSSL methods. CCS CONCEPTS • Computing methodologies → Knowledge representation and reasoning.