Haider Alwasiti

Goal: Building a DL model that can be trained on small EEG training set of a single subject presents an interesting challenge that this work is trying to address. In particular, this study is trying to avoid the need for long EEG data collection sessions, and without combining multiple subjects training datasets, which has a detrimental effect on the classification performance due to the inter-individual variability among subjects. Methods: A customized Convolutional Neural Network with mixup augmentation was trained with $\scriptstyle \mathtt {\sim }$ 120 EEG trials for only one subject per model. Results: Modified ResNet18 and DenseNet121 models with mixup augmentation achieved 0.920 (95% Confidence Interval: 0.908, 0.933) and 0.933 (95% Confidence Interval: 0.922, 0.945) classification accuracy, respectively. Conclusions: We show that the designed classifiers resulted in a higher classification performance in comparison to other DL classifiers of previous studies on the same dataset, despite the limited training dataset used in this work.

Deep metric learning (DML) has achieved state-of-the-art results in several deep learning applications. However, this type of deep learning models has not been tested on the classification of electrical brain waves (EEG) for brain computer interface (BCI) applications. For the first time, we propose a triplet network to classify motor imagery (MI) EEG signals. Stockwell Transform has been used for converting the EEG signals in the time domain into the frequency domain, which resulted in improved DML classification accuracy in comparison to DML with Short Term Fourier Transform (0.647 vs. 0.431). DML model was trained with a topogram of concatenated 64 EEG channel spectrograms. The training batch was comprised of triplet pairs of the anchor, positive, and negative labeled epochs. The triplet network was able to train an embedding feature space that minimized the Euclidean distance between the embeddings of spectrograms of the same class and increased the distance between the embeddings of different labeled images. The proposed method has been tested on an EEG dataset of 109 untrained subjects. We showed that the DML classifier is able to converge with an extremely small number of training samples (~120 EEG trials) for only one subject per model, mitigating the well-known issue of the large inter-individual variability of human MI-BCI EEG which degrades the classification performance. The proposed preprocessing pipeline and the Triplet Network provide a promising method to classify MI-BCI EEG signals with much less training samples than the previous methods.