Graph structure based data augmentation method.

Area of Science:

• Medical signal processing
• Graph-based machine learning
• Data augmentation techniques

Background:

• Medical waveform data, such as electrocardiograms (ECGs), can suffer from inaccuracies due to angular perturbations in lead placement during recording.
• These perturbations negatively impact the performance of algorithmic prediction tasks.
• Existing data augmentation methods may not fully address the structural complexities of waveform data.

Purpose of the Study:

• To propose a novel graph-based data augmentation method for medical waveform data.
• To improve the accuracy and robustness of machine learning models in medical prediction tasks.
• To demonstrate the generalizability of the proposed method across different tasks, models, and datasets.

Main Methods:

• Developed a graph-based data augmentation technique that leverages the inherent graph structures within medical waveform data.
• Applied the method to datasets with known angular perturbations.
• Evaluated the method's performance on prediction tasks and against adversarial attacks.

Main Results:

• Achieved a 1.44% improvement in F1 score across various tasks, models, and datasets.
• Demonstrated enhanced model robustness when tested against adversarial attacks.
• Showed that the proposed Graph Augmentation method can be combined with existing techniques for a further 2.47% gain in F1 score.

Conclusions:

• The proposed graph-based data augmentation method effectively improves the performance and robustness of models using medical waveform data.
• The method is orthogonal to existing augmentation techniques, allowing for synergistic application.
• This approach offers new avenues for data augmentation in medical signal processing.