|
KMID : 1160420210030020099
|
|
Epilia: Epilepsy Commun
2021 Volume.3 No. 2 p.99 ~ p.106
|
|
|
Machine Learning-Based Prediction of Drug-Resistant Epilepsy
|
|
Shin Yong-Won
Eum Hee-Sang
Cha Kwang-Su
Jung Ki-Young
|
|
|
Abstract
|
|
|
Background: A significant proportion of patients with epilepsy suffer from seizures unresponsive to antiepileptic drugs. Certain clinical findings are associated with an increased risk of drug-resistant epilepsy. However, trials of two or more antiepileptic drugs with observational periods are required before diagnosing a patient with drug-resistant epilepsy. Electroencephalography (EEG) is one of the most important tests for epilepsy assessment; however, a visual analysis of EEG usually does not provide evidence regarding whether epilepsy is refractory. By harnessing the capability of machine learning algorithms that recognize and learn patterns that are difficult to recognize by human eyes, we performed a machine learning study of EEG data to predict drug-resistant epilepsy.
Methods: From 2014 to 2017, we collected EEG data from patients diagnosed with epilepsy and used a channel-wise attention network, a convolutional neural network, and a temporal convolutional network to learn the EEG patterns of patients with drug-resistant epilepsy.
Results: Data from 978 EEG examinations were available for training and testing. The best performance achieved was an accuracy of 0.660 and an area under the curve score of 0.634. Our models predicted drug-resistant epilepsy better than drug-sensitive epilepsy.
Conclusions: This result suggests that EEG may contain information predictive of drug-resistant epilepsy; however, the performance of the current model was insufficient for clinical use to predict drug-resistant epilepsy. Our findings warrant further investigation to identify EEG markers of drug-resistance and to increase model performance to a level sufficient to aid in clinical decision-making.
|
|
|
KEYWORD
|
|
|
epilepsy, drug-resistant, machine learning, electroencephalography, prediction
|
|
|
FullTexts / Linksout information
|
|
|
|
|
|
Listed journal information
|
|
|
|