Curriculum Learning using Real and Simulated Data in Deep Learning Models for Electrocardiography Classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Healthcare datasets often lack balance, with a prevalence of healthy subjects, hindering model training.
• Synthetic data generation is explored as a solution for class imbalance, but its efficacy in improving model performance requires validation.
• Electrocardiography (ECG) classification models face challenges due to imbalanced datasets, particularly for rare cardiac conditions.

Purpose of the Study:

• To evaluate the impact of supplementing real ECG data with synthetic ECG signals on the performance of deep neural networks for classification.
• To assess the effectiveness of synthetic data in addressing class imbalance in ECG datasets.
• To investigate the utility of synthetic ECG data for improving the detection of cardiac conduction abnormalities.

Main Methods:

• Utilized the public PTB-XL dataset for real ECG data and MedalCare-XL for generating synthetic ECG signals.
• Employed deep neural networks for classifying four ECG classes, including three cardiac conduction abnormalities.
• Implemented oversampling and data shuffling strategies inspired by Curriculum Learning to optimize training.

Main Results:

• Supplementing real ECG data with synthetic data resulted in a consistent accuracy improvement of up to 0.7%.
• Despite not fully replicating the complexity of real-world ECGs, synthetic data demonstrated a positive impact on classification performance.
• The study confirmed the potential of synthetic ECG data to balance class distributions, especially for rare diseases.

Conclusions:

• Synthetic ECG data can serve as a valuable supplement to real-world data for improving deep learning model performance in ECG classification.
• The findings suggest that synthetic data is a viable strategy for mitigating class imbalance issues in healthcare datasets.
• Further research into synthetic data generation techniques could enhance its complexity and clinical applicability for rare cardiac conditions.