Comparing 1D CNN, LSTM, GRU and Transformer models for PTB-XL ECG classification

Summary

The study compares five deep learning architectures—1D CNN, LSTM, GRU, CNN-LSTM, and Transformer—for multi-label classification of ECG super-classes using the PTB-XL dataset.

Why it matters

Accurate automatic ECG interpretation can reduce dependence on expert clinicians and speed up the diagnosis of cardiovascular diseases. Demonstrating that a lightweight 1D CNN can outperform more complex models suggests a practical solution for resource-constrained settings. The work also adds to the growing body of evidence supporting the applicability of deep learning in health analytics.

Highlights

• The 1D CNN achieved the highest performance after 30 epochs, with a macro F1 score of 0.718 and a macro ROC-AUC of 0.916.
• Convolutional models excel at capturing local waveform patterns, offering a lightweight solution for ECG classification.
• The study treats ECG diagnosis as a multi-label problem across five diagnostic super-classes.

Critical take

The evaluation is limited to a single public dataset (PTB-XL) without external validation, raising questions about the models' generalizability to different clinical environments.