Authors:
Maleyka Seyidova
1
;
2
;
Jasmin Henze
1
;
Arne Pelzer
3
and
Beate Rhein
2
Affiliations:
1
Department of Healthcare, Fraunhofer Institute for Software and Systems Engineering (ISST), Speicherstraße 6, 44147 Dortmund, Germany
;
2
Institute of Computer and Communication Technology (ICCT), Faculty of Information, Media and Electrical Engineering, TH Köln - Cologne University of Applied Sciences, Betzdorfer Straße 2, 50679 Cologne, Germany
;
3
Branch for Hearing, Speech and Audio Technology HSA, Fraunhofer Institute for Digital Media Technology (IDMT), Marie-Curie-Straße 2, 26129, Oldenburg, Germany
Keyword(s):
Data Augmentation, Epileptic Seizure Detection, Imbalanced Data, Convolutional Neural Networks.
Abstract:
Epilepsy, characterized by recurrent seizures, poses a significant risk to an individual’s safety. To mitigate these risks, one approach is to use automated seizure detection systems based on Convolutional Neural Networks (CNN), which rely on large amounts of data to train effectively. However, real-world seizure data acquisition is challenging due to the short and infrequent nature of seizures, resulting in a data imbalance which complicates accurate seizure detection. In this paper, various data augmentation techniques were utilized to increase the amount of training data for CNN, aiming to investigate the potential of these techniques to enhance the performance of the seizure detection algorithm by providing more seizure data. For this purpose, two datasets, a unimodal (3D acceleration) and a multimodal dataset (3D acceleration, heart rate and temperature), were used. To evaluate the effect of the different augmentation techniques, a CNN trained without augmented data was used as
a baseline. Experiments showed that data augmentation techniques improved the seizure detection by lowering the baseline’s false alarm rate while maintaining its high sensitivity. The best results were achieved with a combination of Rotation and Permutation in the multimodal dataset and Rotation, as well as Magnitude Warping, in the unimodal dataset.
(More)