BNCI stands for “Brain and Neural Computer Interface” BNCI adds the “Neural” component to the classic BCI (Brain Computer Interface) with the aim of using other outputs of the Central Nervous System than brain to improve reliability of the interfaces.

In the ABC scope the term BNCI incorporate brain signals analysis adapted to the characteristics of DCP population (the BCI approach). It also includes the use of physiological signals to allow emotion and health management.  Besides the system includes the use of recording movement with inertial measurements units (IMUs) for control and health purposes, both as an activity monitor and as a head posture rehabilitation device.

The BCI component

The effort in the development of the BCI component has been devoted to three different aspects of the EEG signal processing:

  1. Development of artifact reduction algorithms customized to the characteristics of DCP people.
  2. Adaptation and assessment of BCI paradigms for the target users.
  3. Test of new classification approaches to improve classification rates.
Artifact reduction

One of the characteristics of DCP people are involuntary movements. These movements can be more or less continuous in some cases or suddenly triggered by an external factor. In any of these cases, this represents a challenge for any BCI algorithms as movements are one of the most important causes of contamination of the EEG signal.

To tackle with this problem, a new automatic artifact reduction method has been developped, named FORCe [7]. Offline application of the method to EEG recorded from CP users during attempted control of an BCI interface revealed a significant reduction in visually identifiable artifacts, improvement in signal quality index (SQI), increase in ERD/S strength, and increase in classification accuracy. The method is also demonstrated to operate online during EEG acquisition. The FORCe method was compared to other state-of-the-art methods and produced significantly better performance as measured by all the metrics employed. The method has been published OpenAccess. Source code is also, openly available.

A parallel approach for artifact reduction has been crowd-based EEG scoring, discussed with the BCI community at the 6th International BCI Conference in Graz, Austria. Evaluation of artifact reduction is challenging, as there is no “gold standard” performance evaluation. To establish such a standard, a web-based platform has been designed and discussed with other researchers [8], in order to  allows experts to score and annotate EEG signals.

Assessment of BCI paradigms in DCP users

Neural correlates of movement have been investigated, including event-related desynchronization (ERD), phase synchrony, and a recently introduced measure of phase dynamics, in BCI users with CP and healthy control participants. Although present, significantly less ERD (Figure 5) and phase locking were found in the group with CP. Additionally, inter-group differences in phase dynamics were also significant. In short, users with CP exhibit lower levels of motor cortex activation during motor imagery. A paper describing these findings has been published in an open-access peer-reviewed journal [9].

Figura5

Figure 5. Mean band-power differences (left plot) and mean phase locking value differences (right plot), respectively, from healthy users (red colored line) and users with CP (blue) in the mid beta frequency band (16-20Hz) during hand motor imagery. The error bars represent +/- one standard deviation.

Offline within-day and between-day analysis of classification performance in 14 subjects revealed that the combination of connectivity and band power features does not lead to a significant increase in classification performance when using standard linear classifiers.

SCoT is a python toolbox for EEG analysis [10], [11], which allows estimating connectivity between cortical sources that are reconstructed from EEG. The toolbox implements routines for blind source decomposition and connectivity estimation with the MVARICA approach. Additionally, a novel extension called CSPVARICA is available for labeled data. SCoT estimates connectivity from various spectral measures relying on vector autoregressive (VAR) models. Optionally, these VAR models can be regularized to facilitate ill posed applications such as single-trial fitting. While SCoT was mainly designed for application in BCIs, it contains useful tools for other areas of neuroscience. SCoT is a software package that (i) brings combined source decomposition and connectivity estimation to the open Python platform, and (ii) offers tools for single-trial connectivity estimation. Code is released under the MIT license and is available online at github.com/SCoT-dev/ScoT.

Improvement of classification rates

New classification algorithms have been analysed to foresee their performance as BCI classifiers. Random Forest classifiers (RF) have been investigating [12]–[14] for sensor motor rhythm-based BCIs. A systematic analysis of parameterization, online applicability, performance and suitability for online co-adaptation has been carried out. Figure 7 shows the results of an online co-adaptive BCI training procedure. The developed BCI automatically adapts its parameters (filter-bank CSP method and RF classifier) to fit the users EEG patterns (cue-guided paradigm). Peak accuracy computed over 12 users of 85% for binary classification was achieved. This is 10% higher than previous classification systems. Simplicity of handling, speed and robustness of classification (significant higher accuracy compared to linear classifiers) suggest that RF should become the new state-of-the-art classifier for BCI. Results of the studies were presented at a peer-reviewed conference (open access) and submitted for publication to peer-reviewed journal/conference.

Figura6

Figure 6. Wireless version of the IMU (a), a user with the head mounted device driving the PALMIVER vehicle (b).

Figura7

Figure 7. Co-Adaptive 2-class BCI training paradigm.: Mean accuracy over 12 healthy user (140 feedback trials/user). Average (bold red line) and standard deviation (blue dashed line) over all users are shown together with the average performance of individual users (gray line, ‘x’ mark shows peak accuracy). Cue information was provided at t=3s. Feedback was presented from t=3-8s (grey box).

 

Inertial Measurement Units (IMUs) are a small package of sensors, including magnetometers, accelerometers and gyroscopes. Theses sensors allow the estimation of the orientation of the sensor in 3Dspace, and subsequently they are useful as activity monitor of a part of the body.

Within the ABC project, IMUs are attached to the head and allow using the head position as an input device for the communicator. In order to make the IMU more usable, a head-mounted interface has been developed, to control the computer with head movements (Figure 6). The final version is wireless and low-cost.

This device enables users with CP to control the computer and, at the same time, therapists can measure the cervical range of motion, which is useful to detect pathological signs. The head-mounted interface has been tested with the ABC communicator with positive results. Additionally, this device has been integrated into other systems, as the robotic vehicle PALMIBER, designed for cognitive rehabilitation of children with CP. Tests with the vehicle demonstrated that users with CP, who were unable to control the vehicle with the conventional means, were able to drive the vehicle using the head-mounted interface.

Figura8

Figure 8: Architecture of the ABC emotional validation framework.