Reliability and Adaptive Assistance in Brain-Computer Interfaces

Brain-computer interfaces (BCIs) aim at offering an interaction modality for people with severe motor disabilities. Despite promising advances, BCIs are still confronted with multiple challenges in determining user's intentions reliably, mainly due to high performance variations among and within subjects. This issue is usually more critical for end-users with motor impairments. Therefore, in order for BCIs to be used reliably for extended periods of time, they must be able to adapt to the users’ evolving needs. This adaptation should not only be a function of the environmental (external) context, but also should consider the internal context, such as the user's cognitive states and brain signal reliability. The latter, however, has not been extensively studied so far. The objective of this thesis is to investigate possible methods for assessing reliability of BCI and the feasibility of providing adaptive assistance accordingly in a shared control framework. In the first part of the thesis, we propose a taxonomy of shared control frameworks that have been used for BCI applications and we review recently published results in the light of three context-awareness frameworks. Then, following the goal of providing adaptive assistance, we propose two different methods for assessing the reliability of sensorimotor rhythm Electroencephalogram (EEG)-based BCIs by predicting the Command Delivery Time (CDT). The first method involves an information-theoretical approach applied on the EEG signal in the beginning of the mental task execution. The results suggest that it can predict whether a trial is short or long based on the signal in a short period of time (AUC>0.7). The second method is based on the characteristics of the features used by the BCI decoder. As in the previous case, using only a few samples at the beginning of a trial, we are able to predict whether the current trial will be short or long with high accuracies (70%-86%). Both methods resulted in a reliable prediction of CDT, early enough to provide assistance if a command is not reliable. However, the latter was superior, as this prediction could be performed earlier. Therefore, the second method was used to investigate the feasibility of providing online adaptive assistance on a trial-by-trial basis for nine able-bodied subjects. In this experiment, the prediction of short or long trials was used to modulate the command delivery timeout in a BCI game. Results showed significant improvement both in terms of the success rate and the users’ acceptance (assessed using NASA-TLX), when providing adaptive assistance compared to using a fixed timeout. Remarkably, similar results were achieved for an end-user with incomplete locked-in syndrome in a longitudinal study. In addition, consistent BCI performance was obtained over 11 months without re-calibration of the motor-imagery decoder. The results also confirmed both the possibility of a reliable prediction of the CDT (AUC≈0.8) and a significant improvement of the success rate compared to the use of a fixed timeout. Importantly, the online adaptive assistance through the performance estimation was shown to be stable over several sessions of the experiment. The contributions made in this thesis could make a step forward towards improving the usability of BCIs for extended periods of time.


Related material