Author: Tien Dat Nguyen
Nguyen, Tien Dat, 2020 Exploring machine learning architectures for brain computer interface systems, Flinders University, College of Science and Engineering
Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.
Objectives
This study explored the architecture of a machine learning application in a Brain Computer Interface (BCI) system and experimented various classification algorithms in a P300 Speller BCI.
Methods
This study used the P300 Speller BCI datasets from the BCI Competition. The datasets were pre- processed and segmented for analysing and fitting purpose. Wyrm Python toolbox and Scikit-learn Python package were used to help with the exploration. Various classifiers were also experimented: Linear Discriminant Analysis (LDA), Gradient Boosting, Ada Boost, K-neighbours, Decision Tree, Random Forest, Quadratic Discriminant Analysis (QDA), Linear Support Vector Classifier (SVC) and SVC. The predicted characters were compared with true labels downloaded from BCI Competition test results to calculate the percentage of correctly predicted characters.
Results
Main components of a machine learning architecture in a BCI system were explored: EEG signals pre-processing: artifacts removal, features selection, features extraction and predictive model training and results analysing.
For the P300 Speller datasets explored in this study, LDA overall produced better results, with the best score at 94.5% while other classifiers were in a much lower range: <30%, except for the case of Linear SVC with 56% for subject A and 90% for subject B. All those results were observed without Principal component analysis (PCA) and used 64 channels. However, the accuracy dropped when testing with 8 channels, single channel and when testing with fewer trials.
Conclusions
Even though this study only explored machine learning architecture in Speller BCI systems, the knowledge and methodologies can be used to explore P300 BCIs in general. Some classification algorithms were tested, and among those, LDA from Wyrm Python toolbox produced the highest prediction score. However, future work is needed to improve accuracy when using fewer channels or fewer trials.
Keywords: EEG, BCI, brain-computer interface, machine learning, P300
Subject: Computer Science thesis
Thesis type: Masters
Completed: 2020
School: College of Science and Engineering
Supervisor: Trent Lewis