Related Experiment Videos
Sampling Frequency Has Architecture-Dependent Effects on ECG Deep Learning Classification: CNNs Are Robust,
Cole Freedman1, Syed Zawahir Hassan2, Maximo Alvarez1
1School of Medicine, The University of Texas Rio Grande Valley, Edinburg, TX. USA.
Heart Rhythm
|August 4, 2026
Summary
Electrocardiogram (ECG) sampling frequency has minor, architecture-dependent impacts on deep learning classification. Convolutional Neural Networks (CNNs) performed consistently, while Patch Transformers showed slight gains at higher frequencies, increasing computational costs.
Area of Science:
- Biomedical Engineering
- Artificial Intelligence in Medicine
- Signal Processing
Background:
- Electrocardiograms (ECGs) are utilized in diverse settings, with varying sampling frequencies impacting resource utilization.
- Optimizing ECG sampling is crucial for efficient data management and device longevity.
Purpose of the Study:
- To assess how ECG sampling frequency affects classification accuracy and computational demands for CNN and Patch Transformer models.
- To provide insights into selecting appropriate sampling rates for ECG analysis.
Main Methods:
- Utilized the PTB-XL dataset comprising 15,000 12-lead ECGs sampled at 100, 300, and 500 Hz.
- Trained a CNN and a Patch Transformer for multilabel classification of five ECG diagnostic superclasses.
- Evaluated model performance, training duration, memory consumption, and inference speed.
Main Results:
- CNNs demonstrated stable performance across frequencies (macro ROC-AUC ~0.88-0.90).
- Patch Transformers showed incremental performance improvements with higher sampling rates (macro ROC-AUC increased from 0.856 to 0.877).
- Increased sampling frequencies led to higher computational costs (training time, memory, inference latency) for both architectures.
Conclusions:
- ECG sampling frequency has a modest, architecture-specific influence on deep learning classification.
- CNNs are robust to variations in sampling frequency (100-500 Hz).
- Patch Transformers benefit from higher sampling rates but incur greater computational expenses, suggesting a need for task- and architecture-specific sampling strategies in resource-limited environments.