Generalized ECG Heartbeat Classification Using Time-Series Transformers

Abstract

This research investigates Time-Series Transformer architectures for Electrocardiogram (ECG) heartbeat classification, particularly focusing on their generalization capabilities towards new patients and varying signal sampling rates, a critical challenge in real-world clinical applications. This study conducts a systematic comparison between Transformers and Convolutional Neural Network (CNN) models using the St. Petersburg Institute of Cardiological Technics 12-lead Arrhythmia Database (INCART). Key aspects explored include the impact of different input modalities (raw ECG, Continuous Wavelet Transform (CWT) scalograms, and their combination), various Positional Encoding (PE) schemes for Transformers, and the effect of integrating expert-derived RR interval features through different feature fusion techniques. Transformers, especially with concatenation-based PE schemes and CWT or combined ECG+CWT inputs, consistently outperformed CNNs in classification accuracy and generalization when expert features were not used. They demonstrated more than 30% better generalization to unseen patients, and 20% better generalization to unseen patients and sampling rates. Ultimately, this study emphasizes that robust ECG classifiers depend heavily on deliberate architectural choices, positional encoding schemes, input representations, and the integration of expert features to handle inter-patient and sampling rate variations. Importantly, it also demonstrates that Time-Series Transformers can achieve strong results even with relatively modest model sizes and datasets.