Abstract
Epilepsy, a neurological condition characterised by spontaneous and recurring seizures, presents challenges in diagnosis due to the time-consuming nature and requirement for specialised expertise in visually analysing and interpreting electroencephalography (EEG)signals. Therefore, there is a demand for efficient automated methods for seizure type classification.This thesis addresses the intricacies of seizure type classification using EEG signals and utilises the Temple University Hospital EEG Seizure Corpus (TUSZ) v1.5.2 dataset. One notable approach introduced is the Dominant Frequency-based EEG Channel Mapping (DFC-MAP) method, which presents a novel frequency-based dimensionality reduction technique. Compared to traditional methods, DFC-MAP demonstrates superior performance in Ensemble Bagged Trees (EBT) classification (93.88% accuracy) in classifying seven seizure types.This study also explored rebalancing techniques to handle asynchronously distributed data. It was found that the Synthetic Minority Over Sampling Technique (SMOTE) in combination with AdaBoost.M2 yielded the highest accuracy of 99.53% in classifying five seizure types. Furthermore, the thesis investigates the integration of an electrocardiogram (ECG)channel alongside EEG to classify seven seizure types. It was found that the Wavelet Packet Decomposition (WPD) method performed the best (91.46% accuracy) using AdaBoost.M2. Additionally, a comparison between machine learning and deep learning approaches revealed that Long Short-Term Memory (LSTM) marginally outperformedAdaBoost.M2.The efficacy of Mel-frequency Cepstral Coefficients (MFCCs) for feature extraction was also evaluated, alongside the examination of two montage systems: unipolar and bipolar. Results showed that the bipolar montage system, specifically in the Linked Ear(LE) subdirectory of the TUSZ, achieved the highest accuracies for three-class (99.52%) and five-class (99.49%) classifications. Utilising all subdirectories resulted in 99.19%accuracy to classify seven seizure types.In summary, this thesis presents innovative scientific techniques for seizure type classification using EEG signals. The results demonstrate promising advancements, showcasing the potential of automated methods in improving the efficiency and accuracy of epilepsy diagnosis.Thesis is embargoed until 31st March 2026
Date of Award | Mar 2024 |
---|---|
Original language | English |
Supervisor | Mark Ng (Supervisor) & Jim McLaughlin (Supervisor) |
Keywords
- Biomedical signal processing
- Electroencephalogram (EEG)
- Machine learning