Autoencoder error: a new feature for seizure detection

カテゴリ: 研究会(論文単位)

論文No: MBE17018

グループ名: 【C】電子・情報・システム部門 医用・生体工学研究会

発行日: 2017/03/20

タイトル(英語): Autoencoder error: a new feature for seizure detection

著者名: Ziaratnia Sayyed Ali(東京大学),松尾 健(NTT東日本関東病院),川合 謙介(自治医科大学),高橋 宏知(東京大学)

著者名(英語): Sayyed Ali Ziaratnia(The University of Tokyo),Takeshi Matsuo(NTT Medical Center Tokyo),Kensuke Kawai(Jichi Medical University),Hirokazu Takahashi(The University of Tokyo)

キーワード: Epilepsy|Seizure|Diagnose|EEG|Autoencoder

要約(日本語): Typical diagnoses of epilepsy based on long-term electroencephalogram (EEG) are time consuming, and thereby, automatic seizure detection system can be a helpful tool to diagnoses EEG for epileptologists. Conventional seizure detection was based on manually created features, none of which can be applicable for all of patients. In this study, hypothesizing that multi-channel EEG signals are compressible due to spatio-temporal coupling in a state-dependent manner, i.e., seizure or non-seizure, we attempted to use autoencoder (AE) error for seizure detection. Consequently, the AE error was able to classify the seizure and non-seizure states with an accuracy of 85% or more in our data, suggesting that AE error is a candidate of a universal feature for seizure detection.

要約(英語): Typical diagnoses of epilepsy based on long-term electroencephalogram (EEG) are time consuming, and thereby, automatic seizure detection system can be a helpful tool to diagnoses EEG for epileptologists. Conventional seizure detection was based on manually created features, none of which can be applicable for all of patients. In this study, hypothesizing that multi-channel EEG signals are compressible due to spatio-temporal coupling in a state-dependent manner, i.e., seizure or non-seizure, we attempted to use autoencoder (AE) error for seizure detection. Consequently, the AE error was able to classify the seizure and non-seizure states with an accuracy of 85% or more in our data, suggesting that AE error is a candidate of a universal feature for seizure detection.

原稿種別: 英語

PDFファイルサイズ: 1,492 Kバイト