A Comparative Study of Deep Learning Architectures for Multi-Label Electrocardiogram Classification

Abstract

This paper presents a controlled comparative evaluation of convolutional, recurrent, and transformer-based deep learning architectures for multi-label ECG classification using the PTB-XL dataset. Unlike prior studies that vary preprocessing and training regimes across models, all architectures are trained under identical conditions, enabling a fair, controlled assessment of architectural trade-offs that isolates genuine model behavior from preprocessing and optimization artifacts. By training all architectures under identical conditions, this study disentangles genuine architectural advantages from the experimental variability that has obscured conclusions in prior ECG benchmarks. As a result, readers gain actionable insight into which model families are inherently more robust, stable, and clinically viable.