Epileptic seizure detection and prediction based on EEG signal
Zhang, Renjie (2021-09-03)
Epileptic seizure detection and prediction based on EEG signal
(03.09.2021)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe202201178840
https://urn.fi/URN:NBN:fi-fe202201178840
Tiivistelmä
Epilepsy is a kind of chronic brain disfunction, manifesting as recurrent seizures which is caused by sudden and excessive discharge of neurons. Electroencephalogram (EEG) recordings is regarded as the golden standard for clinical diagnosis of epilepsy disease. The diagnosis of epilepsy disease by professional doctors clinically is time-consuming. With the help artificial intelligence algorithms, the task of automatic epileptic seizure detection and prediction is called a research hotspot.
The thesis mainly contributes to propose a solution to overfitting problem of EEG signal in deep learning and a method of multiple channels fusion for EEG features. The result of proposed method achieves outstanding performance in seizure detection task and seizure prediction task.
In seizure detection task, this paper mainly explores the effect of the deep learning in small data size. This thesis designs a hybrid model of CNN and SVM for epilepsy detection compared with end-to-end classification by deep learning. Another technique for overfitting is new EEG signal generation based on decomposition and recombination of EEG in time-frequency domain. It achieved a classification accuracy of 98.8%, a specificity of 98.9% and a sensitivity of 98.4% on the classic Bonn EEG data.
In seizure prediction task, this paper proposes a feature fusion method for multi-channel EEG signals. We extract a three-order tensor feature in temporal, spectral and spatial domain. UMLDA is a tensor-to-vector projection method, which ensures minimal redundancy between feature dimensions. An excellent experimental result was finally obtained, including an average accuracy of 95%, 94% F1-measure and 90% Kappa index.
The thesis mainly contributes to propose a solution to overfitting problem of EEG signal in deep learning and a method of multiple channels fusion for EEG features. The result of proposed method achieves outstanding performance in seizure detection task and seizure prediction task.
In seizure detection task, this paper mainly explores the effect of the deep learning in small data size. This thesis designs a hybrid model of CNN and SVM for epilepsy detection compared with end-to-end classification by deep learning. Another technique for overfitting is new EEG signal generation based on decomposition and recombination of EEG in time-frequency domain. It achieved a classification accuracy of 98.8%, a specificity of 98.9% and a sensitivity of 98.4% on the classic Bonn EEG data.
In seizure prediction task, this paper proposes a feature fusion method for multi-channel EEG signals. We extract a three-order tensor feature in temporal, spectral and spatial domain. UMLDA is a tensor-to-vector projection method, which ensures minimal redundancy between feature dimensions. An excellent experimental result was finally obtained, including an average accuracy of 95%, 94% F1-measure and 90% Kappa index.