Domain randomization using synthetic electrocardiograms for training neural networks.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Wearable single-lead electrocardiogram (ECG) monitors generate vital cardiac data.
• Training robust neural networks for ECG analysis often requires large, diverse, and accurately labeled real-world datasets.
• Challenges include data privacy, class imbalance, and the cost of data acquisition.

Purpose of the Study:

• To develop a method for training neural networks using synthetic ECG data.
• To evaluate the performance of models trained with synthetic data against those trained with real data.
• To demonstrate the effectiveness of domain randomization for improving model generalization.

Main Methods:

• Generation of synthetic ECG signals mimicking wearable single-lead monitors.
• Application of domain randomization, varying waveform shape, RR-intervals, and noise.
• Performance assessment using r-wave detection on real ECGs from various activities and atrial fibrillation.

Main Results:

• Models trained with domain-randomized synthetic ECG data achieved performance on par with or superior to models trained with real data.
• Robust model performance was observed across different seeds and unseen test sets.
• The method demonstrated effective generalization without complex adversarial domain adaptation techniques.

Conclusions:

• Synthetic ECG data generation with domain randomization offers a viable, explainable alternative for training robust neural networks.
• This approach facilitates the use of free, accurately labeled, and private training data.
• It enables domain-insensitive cardiac disease classification and addresses data imbalance issues.