Optimizing Performance of Non-Expert Users in Brain-Computer Interaction by Means of an Adaptive Performance Engine

Brain–Computer Interfaces (BCIs) are become increasingly more available at reduced costs and are being incorporated into immersive virtual environments and video games for serious applications. Most research in BCIs is focused on signal processing techniques and has neglected the interaction aspect of BCIs. This has created an imbalance between BCI classification performance and online control quality of the BCI interaction. This results in user fatigue and loss of interest over time. In the health domain, BCIs provide a new way to overcome motor-related disabilities, promoting functional and structural plasticity in the brain. In order to exploit the advantages of BCIs in neurorehabilitation we need to maximize not only the classification performance of such systems but also engagement and the sense of competence of the user. Therefore, we argue that the primary goal should not be for users to be trained to successfully use a BCI system but to adapt the BCI interaction to each user in order to maximize the level of control on their actions, whatever their performance level is. 

State machine structure. S0 represents the neutral state (indecision). The level of confidence of S-3,3>S-2,2>S-1,1, and W-3,-2,-1,1,2,3 are the state transition thresholds.

To achieve this, we developed the Adaptive Performance Engine (APE) and tested with data from 20 naïve BCI users. BCI-APE is composed by 2 main components: (a) a Bayesian Inference Layer, simpler and more efficient as compared to other supervised learning techniques such as artificial neural networks, in order to formulate the input into a model, where we translate the continuous BCI classification data into probability. As for decision making, we made use of a (b) Finite State Machine because of its efficiency and non-linear properties. APE can provide user specific performance improvements up to approx. 20% and we compare it with previous methods. Finally, we contribute with an open motor-imagery datasets with 2400 trials from naïve users at http://physionet.org/.

Classifier performance comparison, including APE.

Reference:

Ferreira, A., Vourvopoulos, A., & Bermudez I Badia, S. (2015). Optimizing Performance of Non-Expert Users in Brain-Computer Interaction by Means of an Adaptive Performance Engine. In Lecture Notes in Computer Science/Artificial Intelligence (LNCS/LNAI). London, UK: Springer. (Download) (Cite)