Improving generalization performance of electrocardiogram classification models.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models for electrocardiogram (ECG) classification face challenges due to dataset variations.
• Achieving consistent generalization performance across different datasets is crucial for clinical applicability.

Purpose of the Study:

• To develop an improved deep learning model for classifying cardiac abnormalities from 12-lead and reduced-lead ECGs.
• To enhance the generalization performance of ECG classification models across diverse datasets.

Main Methods:

• Employed techniques including constant-weighted cross-entropy loss, additional features, mixup augmentation, squeeze/excitation blocks, and a OneCycle learning rate scheduler.
• Utilized leave-one-dataset-out cross-validation to evaluate generalization performance.
• Applied knowledge distillation from large 12-lead 'teacher' models to improve smaller reduced-lead 'student' models.

Main Results:

• The proposed model achieved high rankings in the PhysioNet Challenge 2021, with scores ranging from 0.55 to 0.58 across different lead configurations.
• Demonstrated superior generalization performance on six hidden test datasets compared to the previous year's submission.
• Knowledge distillation effectively improved the performance of reduced-lead models.

Conclusions:

• The developed deep learning model offers robust cardiac abnormality classification with high generalization capabilities.
• The applied techniques and knowledge distillation are effective strategies for improving ECG analysis models.
• This work advances the potential for reliable AI-driven cardiac diagnostics across varied clinical settings.