Using Brain-Computer Interface (BCI), we can establish a direct communication pathway between the brain and an external device. This thesis focuses on investigating BCI that uses external visual stimuli to elicit steady-state visual evoked potentials (SSVEP) that are captured using electroencephalography (EEG). Through a series of experiments, we study the effect of stimulus parameters, neural signal processing algorithms, and naturalistic stimuli on the speed and classification accuracy of SSVEP-BCI. Our results demonstrate that SSVEP is a viable BCI paradigm that may benefit individuals such as locked-in patients to gain some independence by communicating their commands through the on-screen choices.
We first address the stimulus specificity of BCI. We question how exactly should the visual stimuli be placed spatially, how selective are the frequencies that tag the visual stimuli, what is the effect of stimulus size, and how many visual stimuli can we place simultaneously on the BCI screen in order to elicit the optimal evoked potentials. There is strong evidence that demonstrate the human visual system is of limited processing capability. Hence, addressing the stimulus specificity of BCI allows us to avoid overloading our visual system. An overloaded system tends to generate erratic neural signals that cause poor classification accuracy, which may result in wrong decisions that render VEP-based BCI ineffectual. We conduct our investigations in a multi-part experiment that studies the influence of temporal frequency, stimulus size, and the number of stimuli on the SSVEP-BCI accuracy. Our results show that superior SSVEP-BCI accuracy is attained when stimuli are spatially placed more than 5° apart, with size that subtends at least 2° of visual angle, when using a tagging frequency of between high alpha and beta band (11.3 to 23.3 Hz). These findings may assist in deciding the stimulus parameters for optimal SSVEP-BCI design.
Next, we consider improving the classification performance. Our signal processing routines thus far are based on the power spectral density (PSD) derived from the amplitude of the frequency component present in the Fourier spectrum of the EEG signals. The key benefit of PSD methods is the ease of implementation and effectiveness. However, in the signals with low signal-to-noise ratio, the amplitudes are often masked by substantial noise. Therefore we further research and improve the process by taking advantage of the Fourier phase information. As the amplitude and phase are independent component of the signals, this added dimension enables us to enhance the classification accuracy. The method we use is known as component synchrony measures (CSM). It utilizes the phase synchrony of the transformed signals in the Fourier domain. Our results show a statistical significant improvement in classification accuracy when using a combination of CSM andPSD than using PSD alone.
Prior evidence has shown that stimuli that are more natural and context relevant such as objects, scenes, and faces are better to reveal the functional organization of our visual cortex. We question if such stimuli may be able to better capture user attention and thus improve response time and speed of classification when used in an SSVEP-BCI setup. To study such performance improvements, we need to first establish the effect of using naturalistic visual stimuli on the SSVEP signals. Since the work of Hubel and Wiesel, evidence has shown that there are confounding factors that intervene the responses of the neurons in the visual cortex. We thus developed a framework that guides us in formulating naturalistic stimuli that conform to the selectivity of the visual neurons. The framework made use of the contour, contrast, and edges of the stimuli. It guides us to investigate a balance between context relevance from complex naturalistic stimuli, and maximal evoked responses from elementary stimuli. We parameterize the stimuli based on its edge statistics and seek to elucidate its effect on SSVEP signals. Our results demonstrate that a filtered naturalistic stimulus with a binary contrast level evokes statistically significant SSVEP than a complete naturalistic image. When taking into account the unattended stimuli, our evidence suggests instead that using stimuli of no more than four gray-tone generates the optimal steady-state visual evoked signals.
We next apply the naturalistic stimuli in an open-loop BCI. We are able to obtain comparable classification accuracy to that of the elementary stimuli. Our analysis shows that subjects are able to search and attend to the naturalistic stimuli faster than elementary stimuli. Specifically, there is a statistically significant difference in the speed of classification between elementary stimuli and binary and four-tone type naturalistic stimuli.
Overall, we demonstrate that by accounting the stimuli specificity together with the machine learning algorithm as described above, we are able to derive an SSVEP-BCI that uses naturalistic stimuli with a statistically significant speed enhancement. Our results should contribute towards using SSVEP-BCI as a viable communication tool for patients who lack the normal communication function.