Patterns 2024

DNA_ECG

Decorrelative Network Architecture for Robust Electrocardiogram Classification

Christopher Wiedeman, Ge Wang

Current or former lab member. Select a name to see all of their projects.

Method overview of DNA_ECG
The decorrelation training process (Fig. 6 of the paper).

Abstract

Medical AI needs to recognize when it cannot operate with confidence. Ensembles can estimate uncertainty, but their members often share the same vulnerabilities to adversarial perturbations. This work trains ensemble members to learn different features by combining feature decorrelation with Fourier partitioning, which lowers the chance that one perturbation fools all of them.

The method is tested on single-channel and multi-channel ECG classification, and compared with adversarial training and DVERGE adapted to the same ensemble framework. It generally maintains performance on unperturbed data and shows superior robustness and uncertainty estimation under projected gradient descent and smooth adversarial attacks of various magnitudes. It needs no costly optimization with adversarial samples, so it adds much less computation to training than adversarial training or DVERGE.

What’s in the repository

  • Ensemble training and adversarial evaluation scripts
  • Filters for Fourier partitioning and the feature decorrelation functions
  • Scripts to extract and visualize results; raw ECG inputs are downloaded separately

Citation

@article{wiedeman2024decorrelative,
  title   = {Decorrelative Network Architecture for Robust Electrocardiogram Classification},
  author  = {Wiedeman, Christopher and Wang, Ge},
  journal = {Patterns},
  volume  = {5},
  number  = {12},
  pages   = {101116},
  year    = {2024},
  doi     = {10.1016/j.patter.2024.101116}
}
Back to projects