Klasifikacija EEG signalov z uporabo globokih modelov in pristopov bogatenja podatkov

Eftimska, Iva (author), Dobrišek, Simon (mentor)

Elektroencefalografija (EEG) je eno izmed najpomembnejših orodij za detekcijo in diagnozo epilepsije, saj omogoča neposredno spremljanje električne aktivnosti možganov. EEG je še posebej pomemben zaradi svoje neinvazivnosti in cenovne dostopnosti, kar ga uvršča med široko dostopne diagnostične metode in je primeren za redno uporabo v klinični praksi. V zadnjih letih je v nevroznanosti prisoten hiter razvoj globokega učenja, saj je na voljo čedalje več prostodostopnih podatkovnih zbirk. V tem delu uporabljamo dva globoka modela za klasifikacijo EEG signalov v normalni in abnormalni razred. Prvi je kompleksnejši model ChronoNet, ki doseže 82,12-odstotno točnost pri celotni učni množici, medtem ko drugi, preprostejši ConvNet model, doseže 66,66-odstotno točnost pri celotni učni množici. Z dvema različnima modeloma želimo pokazati, kako delujejo pristopi bogatenja podatkov pri različnih modelih. Modela sta bila naučena na največji dostopni podatkovni zbirki za EEG signale, TUAB zbirki. Pristopi bogatenja podatkov na področjih, kot so računalniški vid in obdelava naravnega jezika, so že uveljavljeni, vendar so pri bioloških signalih, kot je EEG, manj raziskani, ker so EEG signali kompleksni in so izbire primernih pristopov bogatenja podatkov manj intiutivne. Zato bo to delo temeljilo na identifikaciji pravilnega pristopa bogatenja podatkov, pri čemer bi se ohranila struktura signalov. Z našim pristopom smo s povezovanjem Mixup (mešanje primerov samo iz normalnega razreda) in spremembo predznaka dosegli izboljšavo točnosti v primerjavi z izvedbo brez pristopa bogatenja podatkov na 10-odstotni učni množici iz 61,89~\% na 66,09~\%. Izboljšave so bile dosežene tudi pri ostalih metrikah, kot so F1 ocena, natančnost in ROC AUC. Takšne izboljšave so bili statistično značilne, kar smo pokazali z Wilcoxonovim testom, ker smo dosegli p-vrednosti pod 0,05. Tudi za preprostejši model ConvNet so bile dosežene izboljšave, vendar te niso bile statistično značilne.

• Language: Slovenian
• Keywords: EEG, pristopi bogatenja podatkov, binarna klasifikacija, globoko učenje, večji obseg podatkov.
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2024
• PID: 20.500.12556/RUL-164498
• Publication date in RUL: 28.10.2024

Secondary language

• Language: English
• Title: Classification of EEG Signals Using Deep Models and Data Augmentation Approaches

Abstract

Electroencephalography (EEG) is one of the most important tools for the detection and diagnosis of epilepsy, as it allows for the direct monitoring of the brain's electrical activity. EEG is particularly significant due to its non-invasiveness and cost accessibility, placing it among widely available diagnostic methods suitable for regular use in clinical practice. In recent years, the development of deep learning in neuroscience has advanced significantly because increasingly large open-access datasets are available. In this study, we use two deep models to classify EEG signals into normal and abnormal classes. One is the more complex ChronoNet model, which achieves an accuracy of 82,12\% on the entire training dataset, and the other is the simpler ConvNet model, which achieves an accuracy of 66,66\% on the entire training dataset. Two different model capacities were selected to demonstrate how data augmentation approaches work with different models. The models were trained on the largest available EEG signal dataset, the TUAB collection. Data augmentation approaches in fields such as computer vision and natural language processing are already well-established; however, they are less explored for biological signals like EEG because EEG signals are complex and the selection of appropriate data augmentation approaches is less intuitive. Therefore, this work focuses on identifying the appropriate data augmentation approach that preserves the structure of the signal. With our approach, which combines Mixup (mixing samples only within the normal class) and sign flip, we achieved an improvement compared to not using data augmentation, increasing the accuracy on 10\% of the training dataset from 61,89\% to 66,09\%. Improvements were also achieved in other metrics such as F1-score, precision, and ROC AUC. These improvements were statistically significant, as demonstrated by the Wilcoxon test, where we obtained p-values below 0,05. For the simpler ConvNet model, improvements were achieved but were not statistically significant.

• Keywords: EEG, data augmentation approaches, binary classification, deep learning, large-scale dataset.