Sampling frequency has architecture-dependent effects on electrocardiogram deep learning classification:...

Electrocardiograms (ECGs) are deployed across a wide range of environments and frequently operate at varying sampling frequencies, directly influencing bandwidth consumption, storage demands, and battery life.

This study aims to evaluate the impact of ECG sampling frequency on both classification performance and computational cost across convolutional neural network (CNN) and Patch Transformer architectures.

Using the PTB-XL dataset, 15,000 eligible 12-lead ECGs were analyzed at 100, 300, and 500 Hz. A CNN and a Patch Transformer were trained for multilabel classification of 5 diagnostic superclasses: normal, myocardial infarction, ST/T-wave change, conduction disturbance, and hypertrophy. Performance was assessed along with training time, memory use, and inference latency.

CNN performance was relatively stable across sampling frequencies, with macro receiver operating characteristic area under the curve values of 0.895, 0.886, and 0.881 at 100, 300, and 500 Hz, respectively. In contrast, the Patch Transformer showed modest improvement with increasing frequency, with macro receiver operating characteristic area under the curve increasing from 0.856 at 100 Hz to 0.864 at 300 Hz and 0.877 at 500 Hz. Higher sampling frequencies increased computational cost for both architectures, with longer training times, greater memory use, and slower inference. Sensitivity and learning-curve analyses supported architecture-dependent trends across the tested configurations.

ECG sampling frequency had modest, architecture-dependent effects on deep learning classification performance. CNNs were robust across 100-500 Hz, whereas the Patch Transformer benefited incrementally from higher temporal resolution at greater computational cost. These findings support a task- and architecture-aware approach to ECG acquisition in resource-constrained settings.