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

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Deep learning models are increasingly used for automated electrocardiogram (ECG) analysis.
• Preprocessing steps are often assumed to universally improve ECG classification performance.

Purpose of the Study:

• To investigate the impact of signal cleaning, trend removal, and normalization on six deep learning architectures for ECG classification.
• To challenge the conventional assumption that specific preprocessing techniques universally benefit ECG analysis.

Main Methods:

• Evaluation of 24 preprocessing combinations across six leading deep learning architectures (CNNs, Wavelet-based, Transformers).
• Utilized the PTB-XL dataset, a standard reference for ECG data analysis.
• Each trial was repeated ten times to ensure robustness.

Main Results:

• Significant architecture-dependent sensitivity to preprocessing techniques was observed.
• Convolutional neural networks performed best with raw, unnormalized ECG data.
• Wavelet-based models improved with trend removal; Transformers showed broad robustness.

Conclusions:

• Preprocessing pipelines for ECG analysis should be tailored to specific deep learning model architectures.
• The assumption of universally beneficial normalization for bounded signal data in ECG preprocessing is challenged.
• Findings advocate for a nuanced approach to ECG data preprocessing in deep learning.