Classification of imbalanced ECGs through segmentation models and augmented by conditional diffusion model.

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Electrocardiograms (ECGs) are vital for diagnosing arrhythmias, but continuous monitoring is challenged by imbalanced datasets like the MIT-BIH arrhythmia dataset.
• Accurate segmentation of individual heartbeats from continuous ECG data is essential for reliable arrhythmia detection.

Purpose of the Study:

• To develop a robust arrhythmia classification model that overcomes data imbalance issues in the MIT-BIH dataset.
• To compare annotation-based and deep learning-based automated segmentation methods for heartbeats.
• To evaluate the effectiveness of variational autoencoder (VAE) and conditional diffusion for data augmentation.

Main Methods:

• Employed variational autoencoder (VAE) and conditional diffusion for advanced data augmentation to address class imbalance.
• Compared annotation-based segmentation (R-peak labels) with a deep learning-based automated segmentation model.
• Utilized MobileNetV2 architecture for the proposed arrhythmia classification model.

Main Results:

• The proposed model, using MobileNetV2 with annotation-based segmentation and conditional diffusion augmentation, showed a 1.23% F1 score and 1.73% precision improvement over the baseline.
• The model demonstrated accurate classification of a wide range of arrhythmias, including minority classes.
• Achieved enhanced performance compared to models using the original imbalanced dataset.

Conclusions:

• The developed model effectively classifies arrhythmias, including underrepresented classes, by leveraging advanced augmentation and segmentation techniques.
• This research provides a foundation for improved data utilization and model performance in arrhythmia diagnosis.
• The findings contribute to more sophisticated and reliable diagnostic tools, enhancing healthcare services.