Convolutional neural network-based fast seizure detection from video electroencephalograms. (February 2023)

• Record Type: Journal Article
• Authors: Chou, Chi-Hsiang; Shen, Tsu-Wang; Tung, Hsin; Hsieh, Peiyuan F.; Kuo, Chih-En; Chen, Ting-Mao; Yang, Chao-Wei
• Is Part Of: Biomedical signal processing and control. Volume 80:Part 2(2023)
• Publication Date: 2023-02
• Subjects: Computer-aid -- Ensemble -- Proportion of testing dataset -- Video electroencephalogram -- Seizure detection; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104380
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Our study was done on a large seizure dataset (seizure video-EEG duration of 47, 444 s from 30 patients). Our method presented may be a promising computer-aid deep learning approach to detect seizures recorded on video-EEG. For testing phase, selected proportion of testing dataset may influence the performances of deep learning models. The performances of a trained model from one EEG recording method might not be applicable to another EEG recording method. Abstract: Objective: Ictal stage detection in electroencephalography (EEG) is important for epilepsy diagnosis. However, it's a laborious and time-consuming task for neurologists, especially with lengthy EEG recordings (such as video-EEG). Here we applied the current deep learning techniques to improve ictal stage detection efficiency. Methods: We enrolled epileptic patients at video-EEG examination beds of our neurology department between Jan 2016 and Apr 2020. Four stages of EEG (interictal, preictal, ictal and postictal) were labeled and confirmed by neurologists. Labeled EEG signals were transformed into second-order Poincaré difference plots. The whole dataset was divided into 'Training', 'Validation' and 'Testing' datasets with the proportions of 45%, 5% and 50% respectively. We applied 4 Convolutional Neural Networks (CNNs), namely, AlexNet, GoogLeNet, ResNet50 and DenseNet201, to learn classifying. The top three CNNs were put into an ensemble model for datasets classification. Results: 30 epileptic patients were included: Male/Female 16/14, mean age 42.8 ± 15.0 (range 22 ∼ 81) years old. Total recorded ictal duration was 47, 444 s. Overall accuracies for each of the 4 CNN models (AlexNet, GoogLeNet, ResNet50 and DenseNet201) were 78.0%, 78.3%, 81.7% and 80.8%. We omitted the AlexNet, the other 3 CNNs were put into the ensemble model. The best performance was under test with 1% of the Testing dataset yielded the overall accuracies of 85.9%. The ictal stage accuracy was 97.7%. Conclusion: Our method is a promising computer-aid deep learning approach to detect seizures on video EEG. Significance: For testing phase, selected proportion of testing dataset may influence the performances of deep learning models.