Impact of ECG Characteristics on Deep Classifiers

Deep learning has shown great promise in automating electrocardiogram (ECG) analysis, yet its performance varies significantly across different clinical settings and data conditions. This study systematically investigates how deep learning classifiers interact with ECG-related characteristics, as influenced by key data factors including data size, data augmentation, data filtering, signal length, and label correlations. We benchmark two representative architectures, ResNet and Transformer, on four public realworld ECG datasets. Our findings reveal that model performance varies substantially based on data factor configurations, often exceeding the performance gap between different model architectures. Notably, ResNet frequently outperforms Transformer under challenging conditions, not due to superior complexity, but due to greater robustness to data quality issues. These results underscore a critical gap in current ECG AI research: the underappreciation of data-centric optimization. Addressing this gap by integrating model development with informed data strategies is essential for unlocking the full potential of deep learning in real-world ECG analysis.