|Resumo: ||Stroke and road traffic injuries may severely affect movements of lower-limbs in humans, and consequently the locomotion, which plays an important role in daily activities, and the quality of life. Robotic exoskeleton is an emerging alternative, which may be used on patients with motor deficits in the lower extremities to provide motor rehabilitation and gait assistance. However,
the effectiveness of robotic exoskeletons may be reduced by the autonomous ability of the robot to complete the movement without the patient involvement. Then, electroencephalography signals (EEG) have been addressed to design brain-computer interfaces (BCIs), in order to
provide a communication pathway for patients perform a direct control on the exoskeleton using the motor intention, and thus increase their participation during the rehabilitation. Specially, activations related to motor planning may help to improve the close loop between user and exoskeleton, enhancing the cortical neuroplasticity. Motor planning begins before movement onset, thus, the training stage of BCIs may be affected by the intuitive labeling process, as it is not possible to use reference signals, such as goniometer or footswitch, to select those time periods really related to motor planning. Therefore, the gait planning recognition is a challenge,
due to the high uncertainty of selected patterns, However, few BCIs based on unsupervised methods to recognize gait planning/stopping have been explored.
This Doctoral Thesis presents unsupervised methods to improve the performance of BCIs during gait planning/ stopping recognition. At this context, an adaptive spatial filter for on-line processing based on the Concordance Correlation Coefficient (CCC) was addressed to preserve
the useful information on EEG signals, while rejecting neighbor electrodes around the electrode of interest. Here, two methods for electrode selection were proposed. First, both standard deviation and CCC between target electrodes and their correspondent neighbor electrodes are analyzed on sliding windows to select those neighbors that are highly correlated. Second, Zscore analysis is performed to reject those neighbor electrodes whose amplitude values presented significant difference in relation to other neighbors.
Furthermore, another method that uses the representation entropy and the maximal information compression index (MICI) was proposed for feature selection, which may be robust to select patterns, as only it depends on cluster distribution. In addition, a statistical analysis was introduced here to adjust, in the training stage of BCIs, regularized classifiers, such as
support vector machine (SVM) and regularized discriminant analysis (RDA).
Six subjects were adopted to evaluate the performance of different BCIs based on the proposed methods, during gait planning/stopping recognition.
The unsupervised approach for feature selection showed similar performance to other methods based on linear discriminant analysis (LDA), when it was applied in a BCI based on the traditional Weighted Average to recognize gait planning. Additionally, the proposed adaptive filter improved the performance of BCIs based on traditional spatial filters, such as Local Average Reference (LAR) and WAR, as well as others BCIs based on powerful methods, such as Common Spatial Pattern (CSP), Filter Bank Common Spatial Pattern (FBCSP) and Riemannian kernel (RK). RK presented the best performance in comparison to CSP and FBCSP, which
agrees with the hypothesis that unsupervised methods may be more appropriate to analyze clusters of high uncertainty, as those formed by motor planning. BCIs using adaptive filter based on Zscore analysis, with an unsupervised approach for feature selection and RDA showed promising results to recognize both gait planning and gait stopping, achieving for three subjects, good values of true positive rate (>70%) and false positive (<16%). Thus, the proposed methods may be used to obtain an optimized BCI that preserves the useful information, enhancing the gait planning/stopping recognition. In addition, the method for feature selection has low computational cost, which may be suitable for applications that
demand short time of training, such as clinical application time.|