Verbesserung interaktiven Lernens von Mensch und Maschine zur Überwindung der Unfähigkeit eine Gehirn-Computer Schnittstelle zu steuern
Final Report Abstract
This project VitalBCI2 is a followup of the project VitalBCI1. The goal of the whole VitalBCI endeavor was to improve the coadaptive learning process of user and BCI system in order to improve the efficiency of BCI control, in particular to allow a larger population of users an efficient BCI control. Furthermore, we aimed at defining predictors of BCI control. The predecessor project VitalBCI1 elucidated some prominent problems, that prevented a number of potential BCI users to gain control. Moreover, we determined a neurophysiological predictor and two psychological predictors that allow, already before the start of a BCI session, to estimate the accuracy of BCI control. Rather than only seeing these predictors as a value of their own, these were taken as a starting point to better understand the problem of BCI inefficiency, and to devise countermeasures. Accordingly, the followup VitalBCI2 was mainly concerned with deepening and exploiting the knowledge about predictors of BCI performance and, moreover, with the development of adaptive methods for EEG signal processing and classification to foster a coadaptive calibration approach in which BCI systems and their users are evolving in mutual interaction. Our first aim of this followup project was to validate the correlation between the neurophysiological predictor found in the predecessor project VitalBCI1 and BCI performances on a new large pool of participants. We found a significant positive correlation of 0.53 (p<0.01), which rose to 0.66 after outlier rejection, confirming the results of the previous study. Importantly, our results show in addition that the proposed model can be transferred between two SMRBCI studies that employ different designs. Indeed, the BCI performances of new participants could be predicted with high significant correlation (p<0.01) by the prediction model derived from the previous dataset and with a root mean square error (RMSE) of 16%. The replication of the neurophysiological predictor is a major achievement of the project, since the participants tested were different and also the BCI procedure changed with the introduction of the coadaptive calibration approach. These results consolidate the validity of the developed prediction model across different participants and experimental protocols. Our second aim was to find out whether the psychological predictors that we found in the predecessor project to correlate with BCI performance (concentration / relaxation and visuomotor coordination) can also be used as an active intervention to improve BCI performance. Visuomotor coordination performance, the ability to concentrate and the resting state µpeak together explained about 64% of the variance of SMR BCI performance in a large sample of N=40 healthy participants, whereby the contribution of the psychological factors was around 25%. In this study subjects participated in only one session and the respective machine learning approach from the Berlin Brain Computer Interface (BBCI) was applied. We argued that psychological factors may play a more important role when true neurofeedback learning is required. Thus, a study was conducted in which no machine learning was added and subjects were required to regulate the SMR amplitude within 3 sessions. The predictor model for visuomotor coordination determined by Hammer and colleagues (2012) again explained about 12% of the variance in a group of N=33 healthy participants. Thus, in the followup studies we addressed specifically the visuomotor coordination ability and concentration whereby the latter was only indirectly targeted via manipulating the relaxation level. We assumed that participants who are more relaxed would be able to better concentrate on the task. For the analysis of the psychological predictors we had to analyse the Berlin and Würzburg groups separately and, however, results could not fully consolidate the two predictors of SMR BCI performance. Complementary to the main thread that is devoted to improve BCI systems that rely on the voluntary modulation of sensorimotor rhythms (SMR), we explored a different category of BCIs based on eventrelated potentials (ERPs). This parallel endeavor was motivated by the hypothesis that it might not be possible, even with advanced machine learning (ML) methods, to decrease the BCI inefficiency in SMRBCIs to a nonnegligible factor, whereas ERPbased BCIs were expected to have a much larger applicability in the population. Still, a limiting factor of ERPbased BCIs in their applicability to patients was that these systems had experimental paradigms that are effective only under the condition of relatively good oculomotor control, which is not necessarily available in the target user group of, e.g., ALS patients. For this reason we complemented the effort to develop gazeindependent ERP spellers. Finally, we have developed a number of methods for adapting the BCI system to the user, which are integrated in our multilevel approach to coadaptive calibration. This methodology was used in the large scale study, which is at the core of this project. In parallel to that study, we conducted a series of smaller studies which explored further developments in the coadaptive approach. Moreover, we explored novel techniques of machine learning in order to make the BCI classification process more robust.
