Show simple item record

dc.rights.licenseCC-BY-NC-ND
dc.contributor.advisorMaanen, Leendert van
dc.contributor.authorSterk, Hunter
dc.date.accessioned2022-07-15T00:01:15Z
dc.date.available2022-07-15T00:01:15Z
dc.date.issued2022
dc.identifier.urihttps://studenttheses.uu.nl/handle/20.500.12932/41764
dc.description.abstractRecent research on classifying motor imagery electroencephalography (EEG) signals has often shown that exceptionally high accuracy scores can be achieved. However, this is usually done with EEG caps with a lack of usability and by performing the classification on large time windows. When using a brain computer interface (BCI) for real time gaming, it is important to keep the time windows as short as possible to minimise the response time in the game. The aim of this study is to investigate whether the performance of the most common and state of the art methods for classifying motor imagery EEG signals is sufficient for real time BCI gaming. The data is recorded with portable dry electrodes and is split into small windows of 512 ms. Four different classifiers were evaluated: Logistic Regression (LR), Support Vector Machines (SVM), Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM). For the feature extraction, Common Spatial Patterns (CSP), Fast Fourier transform (FFT), Discrete Wavelet Transform (DWT) and statistical features are used. The LSTM model used another feature extraction method, in which each time window was broken down into eight segments and converted to features by applying a linear regression on each segment. The LSTM model outperformed the other models with a maximum accuracy score of 75.1%. The other three classifiers performed best with a combination of CSP, DWT and statistical features, but was not able to exceed a score of 57.6%. When the LSTM model was applied on other datasets, the accuracy score dropped significantly. When using the model for BCI gaming, it is recommended to record the data and train the model right before each gaming session. With the methods used in this study, it was impossible to use a model that has been trained at a different moment or by another person. More research could be done into a technique like transfer learning, in order to reuse already trained models. The conclusion of this study is that the performance achieved by the used methods is indeed much lower when used on data from dry electrodes with small window sizes. With proper mental tasks, it seems that a simple BCI game can be played with the evaluated models. But all the high results from recent studies give a distorted view of the possibilities for BCI gaming.
dc.description.sponsorshipUtrecht University
dc.language.isoEN
dc.subjectA study on the performance of classifying motor imagery electroencephalography (EEG) by using dry electrodes in combination with a classification on large time windows. The main question is whether this performance is sufficient for real time brain computer interface (BCI) gaming.
dc.titleThink & Play: Classifying Left and Right Hand Motor Imagery EEG Signals by Using Dry Electrodes for Real-Time BCI Gaming
dc.type.contentMaster Thesis
dc.rights.accessrightsOpen Access
dc.subject.keywordsbrain computer interface (BCI); motor imagery; electroencephalography (EEG); BCI gaming; dry electrodes; signal processing; response time; window size; feature extraction; classification; deep neural networks; convolutional neural networks; recurrent neural networks
dc.subject.courseuuArtificial Intelligence
dc.thesis.id5530


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record