Graph Structure Based Data Augmentation Method

TL;DR

This work tackles lead-position perturbations in medical waveform predictions by exploiting the intrinsic lead-graph structure. The authors propose a graph-based augmentation that builds a lead-to-lead adjacency $A$ from average cross-lead correlation, computes $\widetilde{X}^{(i)}=\sum_{j\neq i} A_{ij}X^{(j)}$, and forms $\widehat{X}^{(i)}=(1-\lambda)X^{(i)}+\lambda \widetilde{X}^{(i)}$, with random application controlled by $p$ and $\lambda$ drawn from $U(0,\alpha)$, while also incorporating RandAugment-style transformations. Empirical results on three open ECG datasets across three architectures show that graph augmentation yields performance gains comparable to tuned normal augmentations and offers additive improvements when combined, while enhancing robustness to adversarial perturbations. The findings suggest a general, orthogonal augmentation strategy for graph-structured biomedical waveforms and point to potential extension to EEG and other modalities.

Abstract

In this paper, we propose a novel graph-based data augmentation method that can generally be applied to medical waveform data with graph structures. In the process of recording medical waveform data, such as electrocardiogram (ECG) or electroencephalogram (EEG), angular perturbations between the measurement leads exist due to discrepancies in lead positions. The data samples with large angular perturbations often cause inaccuracy in algorithmic prediction tasks. We design a graph-based data augmentation technique that exploits the inherent graph structures within the medical waveform data to improve both performance and robustness. In addition, we show that the performance gain from graph augmentation results from robustness by testing against adversarial attacks. Since the bases of performance gain are orthogonal, the graph augmentation can be used in conjunction with existing data augmentation techniques to further improve the final performance. We believe that our graph augmentation method opens up new possibilities to explore in data augmentation.

Figures (4)

• Figure 1: Comparison between an original signal and its corresponding augmentation
• Figure 2: Overall structure of augmentation module
• Figure 3: Performance of ResNet trained with different augmentation parameter combinations selected by RandAugment
• Figure 4: Adversarial robustness of ResNets trained using RandAugment and using both RangAugment and graph augmentation according to different perturbation strength ($\epsilon$)