UC Irvine

Brain-computer interfaces (BCIs) are a means of controlling devices in our environment using brain signals. They are promising tools for persons with neurological disabilities including speech and muscular impediments that externally connect the brain and affected muscle groups using computational devices. There are several brain imaging modalities categorized as invasive or non-invasive with electroencephalography (EEG) being the most popular due to its practicality and affordability. A BCI pipeline consists of data acquisition, preprocessing, feature extraction, classification followed by an action by a device. There are multiple machine learning (ML) algorithms that can be used for classification however there are no guidelines for selecting an algorithm for a certain control signal or BCI application. In this thesis, we compare five common ML algorithms, namely Logistic regression (LR), neural networks (NNs), support vector machines (SVMs), k-nearest neighbor (KNN) and a Bayesian classifier for two common control signals used in EEG-BCIs, the P300 potential and sensorimotor rhythms. For P300, the Bayesian decoder and a fully connected NN performed the best for cross-validations performed on each session's data. SVM performed the best for leave-one-subject-out cross-validations where the training data didn't include any of the test subject's data. SVM and LR performed the best for sensorimotor data for cross-validations on each session. A procedure mimicking real-time decoding performed on the sensorimotor data didn't differentiate between the algorithms in terms of accuracy however a recurrent neural network based on long short term memory units had the lowest lags defined as the temporal offset between the predictions and true labels. These results provide a good foundation for the selection of ML algorithms for BCI applications. Future works can build on this by incorporating more ML algorithms and testing on additional BCI control signals.