| Popis: |
A Brain-Computer Interface (BCI) is a device that uses the brain activity of a person as an input to select desired outputs on a computer. BCIs that use surface electroencephalogram (EEG) recordings as their input are the least invasive but also suffer from a very low signal-to-noise ratio (SNR) due to the very low amplitude of the person’s brain activity and the presence of many signal artefacts and background noise. This can be compensated for by subjecting the signals to extensive signal processing, and by using stimuli to trigger a large but consistent change in the signal – these changes are called evoked potentials. The method used to stimulate the evoked potential, and introduce an element of conscious selection in order to allow the user’s intent to modify the evoked potential produced, is called the BCI paradigm. However, even with these additions the performance of BCIs used for assistive communication and control is still significantly below that of other assistive solutions, such as keypads or eye-tracking devices. This thesis examines the paradigm and signal processing components of BCIs and puts forward several methods meant to enhance BCIs’ performance and efficiency. Firstly, two novel signal processing methods based on Empirical Mode Decomposition (EMD) were developed and evaluated. EMD is a technique that divides any oscillating signal into groups of frequency harmonics, called Intrinsic Mode Functions (IMFs). Furthermore, by using Takens’ theorem, a single channel of EEG can be converted into a multi-temporal channel signal by transforming the channel into multiple snapshots of its signal content in time using a series of delay vectors. This signal can then be decomposed into IMFs using a multi-channel variation of EMD, called Multi-variate EMD (MEMD), which uses the spatial information from the signal’s neighbouring channels to inform its decomposition. In the case of a multi-temporal channel signal, this allows the temporal dynamics of the signal to be incorporated into the IMFs. This is called Temporal MEMD (T-MEMD). The second signal processing method based on EMD decomposed both the spatial and temporal channels simultaneously, allowing both spatial and temporal dynamics to be incorporated into the resulting IMFs. This is called Spatio-temporal MEMD (ST-MEMD). Both methods were applied to a large pre-recorded Motor Imagery BCI dataset along with EMD and MEMD for comparison. These results were also compared to those from other studies in the literature that had used the same dataset. T-MEMD performed with an average classification accuracy of 70.2%, performing on a par with EMD that had an average classification accuracy of 68.9%. Both ST-MEMD and MEMD outperformed them with ST-MEMD having an average classification accuracy of 73.6%, and MEMD having an average classification accuracy of 75.3%. The methods containing spatial dynamics, i.e. MEMD and ST-MEMD, outperformed those with only temporal dynamics, i.e. EMD and T-MEMD. The two methods with temporal dynamics each performed on a par with the non-temporal method that had the same level of spatial dynamics. This shows that only the presence of spatial dynamics resulted in a performance increase. This was concluded to be because the differences between the classes of motor-imagery are inherently spatial in nature, not temporal. Next a novel BCI paradigm was developed based on the standard Steady-state Somatosensory Evoked Potential (SSSEP) BCI paradigm. This paradigm uses a tactile stimulus applied to the skin at a certain frequency, generating a resonance signal in the brain’s activity. If two stimuli of different frequency are applied, two resonance signals will be present. However, if the user attends one stimulus over the other, its corresponding SSSEP will increase in amplitude. Unfortunately these changes in amplitude can be very minute. To counter this, a stimulus amplitude and frequency of the vibrotactile stimuli. It was hypothesised that if the stimuli generator was constructed that could alter the were of the same frequency, but one’s amplitude was just below the user’s conscious level of perception and the other was above it, the changes in the SSSEP between classes would be the same as those between an SSSEP being generated and neutral EEG, with differences in α activity between the low-amplitude SSSEP and neutral activity due to the differences in the user’s level of concentration from attending the low-amplitude stimulus. The novel SSSEP BCI paradigm performed on a par with the standard paradigm with an average 61.8% classification accuracy over 16 participants, compared to an average 63.3% classification accuracy respectively, indicating that the hypothesis was false. However, the large presence of electro-magnetic interference (EMI) in the EEG recordings may have compromised the data. Many different noise suppression methods were applied to the stimulus device and the data, and whilst the EMI artefacts were reduced in magnitude they were not eliminated completely. Even with the noise the standard SSSEP stimulus paradigm performed on a par with studies that used the same paradigm, indicating that the results may not have been invalidated by the EMI. Overall the thesis shows that motor-imagery signals are inherently spatial in difference, and that the novel methods of T-MEMD and ST-MEMD may yet out-perform the existing methods of EMD and MEMD if applied to signals that are temporal in nature, such as functional Magnetic Resonance Imaging (fMRI). Whilst the novel SSSEP paradigm did not result in an increase in performance, it highlighted the impact of EMI from stimulus equipment on EEG recordings and potentially confirmed that the amplitude of SSEP stimuli is a minor factor in a BCI paradigm. |
