Architecture-Specific Impact of Preprocessing on Machine Learning Models for ECG Classification

Automated analysis of electrocardiograms relies increasingly on deep learning models. In these models, preprocessing steps may often be applied under the assumption that they inherently improve performance. This study investigates the effects of signal cleaning, trend removal, and normalization on six leading deep learning architectures for electrocardiogram classification, based on the well-established PTB-XL dataset, a widely used reference standard in electrocardiogram (ECG) data analysis. By exhaustively evaluating 24 combinations of preprocessing techniques with each model, and repeating each trial ten times for robustness, we find a surprising degree of architecture-dependent sensitivity. Notably, convolutional neural networks performed best when raw, unnormalized electrocardiograms are used, in direct contradiction to prevailing assumptions. In contrast, wavelet-based models benefited substantially from trend removal, while transformer-based models demonstrated robustness across a broad range of preprocessing options. These findings suggest that preprocessing pipelines should be designed with specific model architectures in mind, challenging the conventional wisdom of applying normalization in preprocessing for bounded signal data.