| Brain computer interface systems based on electroencephalograph (EEG) signals have limitations which challenge their application as a practical device for general use. The signal features generated by the brain states we wish to detect possess a high degree of inter-subject and intra-subject variation. Additionally, these features usually exhibit a low variation across each of the target states. Collection of EEG signals using low resolution, non-invasive scalp electrodes further degrades the spatial resolution of these signals. The majority of brain computer interface systems to date require extensive training prior to use by each individual user. The discovery of subject invariant features could reduce or even eliminate individual training requirements. To obtain suitable subject invariant features requires search through a high dimension feature space consisting of combinations of spatial, spectral and temporal features. Poorly separable features can prevent the search from converging to a usable solution as a result of degenerate classifiers. In such instances the system must detect and compensate for degenerate classifier behavior. This dissertation presents a method to accomplish this search using a wrapper architecture comprised of a sequential forward floating search algorithm coupled with a support vector machine classifier. This is successfully achieved by the introduction of a scalar Quality (Q)-factor metric, calculated from the ratio of sensitivity to specificity of the confusion matrix. This method is successfully applied to a multiclass subject independent BCI using 10 untrained subjects performing 4 motor tasks. |
