A study on the utility of temporal derivatives and unsupervised clustering in brain-computer interfaces

Brain-computer interfaces (BCIs) rely on accurate classification of event-related potentials (ERP), a task commonly delegated to a machine-learning algorithm, which investigates features derived from the voltages (V) recorded at different scalp locations with the electro-encephalogram (EEG). The performance of the machine-learning algorithm is an area that has captured the interest of the research community. Although major advancements have been made, BCIs suffer from uncertainties that arise from assumptions such as that participants are “focused”, “still” and that no unpredictable events occurred during the recording, for example abrupt sounds or light changes. From the range of possible uses of BCIs, one of the most challenging is its adaptation to everyday life situations. Addressing both participant and environmental related influences to the EEG could enable the usage of BCIs outside the confines of the laboratory. In addition, in order to create a BCI that can act as an “enhancement” for the able-bodied requires a way to identify recurrent events without prior knowledge, thus providing the user with a way to increment the “understanding” of his BCI. Moreover, information such as location, latency and shape of recurring events could provide solid grounds for future researchers to build upon. In the thesis the above problem is challenged by investigating two main topics: assuming that the neuro-signals are additive (i.e. uncorrelated), (a) the usage of the first time derivative of V (dV) as feature regarding performance in classification of an ERP, and (b) unsupervised clustering of ERPs. Both investigations tackle the problem of mining properties of unknown neuro-signals. Theoretical investigations carried out on in each topic are performed using synthetic signals to assess the expected behaviour. Using real data from a P300 BCI mouse, both topics were evaluated; the classification performance of dV was found to be significantly better than V while evaluating a baseline for comparison. Having such a positive outcome encouraged an attempt to create a single linkage unsupervised clustering method based on statistical significance. Without knowing if an ERP was generated or not, the developed clustering algorithm, based on dV, is shown to be accurate in identifying the shape of the underlying, “unknown” ERP. For years researchers have been constructing experiments to uncover EEG events directly related to stimuli. An outcome of this research is that recurring EEG responses which might have been neglected, simply because they were not expected, are now identifiable.