Feature engineering for performance improvement of p300-based BCIs
Year: 2020
Type: dissertation
Abstract: Good accuracy and Information Transfer Rate (ITR) are required to apply the smart home Brain-Computer Interface (BCI) system that controls home appliances to electroencephalography (EEG) in daily life. To improve the performance, studies on EEG preprocessing, feature extraction, feature selection, and classification model optimization are ongoing. The purpose of this study is to improve performance through feature extraction and sampling in the feature engineering stage of P300-based BCIs. I studied which method is effective among feature extraction methods, whether feature extraction through deep learning algorithm is effective, and how to improve performance by solving data imbalance problem with sampling technique. \nIn this study, I focused on feature extraction and sampling to evaluate the performance of each technique. I used the basic ERP Waveform, morphological features, the features applied to principal component analysis, a convolutional neural network (CNN), a recurrent neural network (RNN), and a stacked auto-encoder (SAE) technique. \nAlso, the odd-ball paradigm used in the smart home BCI system causes data imbalance problems for target stimuli and non-target stimuli. Data imbalance can cause overfitting problems and poor classification performance. I addressed this problem through various sampling techniques. \nThis study used the smart home BCI system user data for 15 subjects for door lock, 15 subjects for electric light, and 14 subjects for Bluetooth speaker. After building a classification model with training data consisting of 50 blocks, I calculated the accuracy and ITR in 30 online controls. \nAs a result, the accuracy of the support vector machine classifier model with the convolutional neural network using a kernel which is sized to account for all the electrodes was the best accuracy (87.13%). However, when over-sampling was applied, the accuracy of the model trained on the classifier with only the Event-Related Potential (ERP) waveform was 87.27%, which was the best accuracy. Especially borderline-SMOTE led to marginally significant performance improvement for poor performers. Also, the models extracted through deep learning were significantly more influenced on training duration than models without feature extraction methods of deep learning. \nThis result suggests that deep learning can be used as feature extraction methods in the P300-based BCIs, and borderline-SMOTE can improve the accuracy and ITR by preventing the classification result from being biased to non-target class. Compared to the previous study, the door lock BCI system had a 5.3% accuracy gain and the electric light BCI system had a 12.67% accuracy gain. In the future, comprehensive performance evaluation research is needed considering not only accuracy and ITR, but also computational cost.
Author: Lee, Tae-Jun
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Keywords: P300, brain-computer interface, feature extraction, deep feature extraction, convolutional neural networks, recurrent neural networks, stacked autoencoder, class imbalance, SMOTE, Borderline-SMOTE, oddball paradigm, event-related potentials, information transfer rate, smart home control, home appliance control, EEG classification, support vector machine, morphological features, principal component analysis, oversampling
Open Access status: green
OpenAlex ID: W6981806064
