Jingna Qiu

Advisors:
Maximlian Oppelt (FIIS), Michael Nissen (M. Sc.), Prof. Dr. Bjoern Eskofier, Dr. med. Lars Anneken (UK Erlangen)

Duration:
11 / 2020 – 05 / 2021

Abstract:
Deep learning models have been successfully used in electrocardiogram (ECG) classification tasks [1]. One potential option to improve the current benchmark is to combat model overfitting. One cause of overfitting is a non-representative training set. The theory of stochastic approximation indicates that a locally optimal network design requires an infinite number of training data for convergence, which, in practice, is not possible. Thus, the network may show a poor generalization ability if the training and validation set are highly apart distributed [2]. Data augmentation is a solution to increase the diversity of training samples such that the distance between the training and validation set is minimized [3].

Image data augmentations have been extensively studied in the computer vision community in order to improve model performance. Many of them, like random cropping and flipping, proved as effective and are commonly used as standard operations for model training. However, for time-
series signals such as ECG curves, the effects of data augmentations are less investigated. Most of the existing validations are based on introducing randomness of digital signals, like oversampling by window slicing [4] or adding some random noises at different frequencies [5].

References:
[1] Zahra Ebrahimi, Mohammad Loni, Masoud Daneshtalab, and Arash Gharehbaghi. A review on deep learning methods for ecg arrhythmia classication. Expert Systems with Applications: X, 7:100033, 2020.
[2] George N Karystinos and Dimitrios A Pados. On overtting, generalization, and randomly expanded training sets. IEEE Transactions on Neural Networks, 11(5):10501057, 2000.
[3] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of Big Data, 6(1):60, 2019.
[4] Zhicheng Cui, Wenlin Chen, and Yixin Chen. Multi-scale convolutional neural networks for time series classication. arXiv preprint arXiv:1603.06995, 2016.
[5] Maximilian P Oppelt, Maximilian Riehl, Felix P Kemeth, and Jan Stean. Combining scatter transform and deep neural networks for multilabel electrocardiogram signal classication. arXiv preprint arXiv:2010.07639, 2020.