Section: New Results
Improving EEG Signal Processing for Brain-Computer Interfaces
Participants: Aurélien Appriou, Satyam Kumar, Fabien Lotte
A review of classification algorithms for BCI: Most current Electroencephalography (EEG)-based Brain-Computer Interfaces (BCIs) are based on machine learning algorithms. We surveyed the BCI and machine learning literature to identify the classification approaches that have been investigated to design BCIs. We found that the recently designed classification algorithms for EEG-based BCIs can be divided into four main categories: adaptive classifiers, matrix and tensor classifiers, transfer learning and deep learning, plus a few other miscellaneous classifiers. Among these, adaptive classifiers were demonstrated to be generally superior to static ones, even with unsupervised adaptation. Transfer learning can also prove useful although the benefits of transfer learning remain unpredictable. Riemannian geometry-based methods have reached state-of-the-art performances on multiple BCI problems and deserve to be explored more thoroughly, along with tensor-based methods. Shrinkage linear discriminant analysis and random forests also appear particularly useful for small training samples settings. On the other hand, deep learning methods have not yet shown convincing and consistent improvement over state-of-the-art BCI methods. This survey was published in Journal of Neural Engineering in .
Exploring Modern Machine Learning Methods to Estimate Mental Workload From EEG Signals: Estimating mental workload from brain signals such as EEG has proven very promising in multiple HCI applications, e.g., to design games or educational applications with adaptive difficulty. However, currently obtained workload classification accuracies are relatively low, making the resulting estimations not fully trustable. We thus studied promising modern machine learning algorithms, including Riemannian geometry-based methods and Convolutional Neural Networks, to estimate workload from EEG signals. We studied them with both user-specific and user-independent calibration, to go towards calibration-free systems. Our results suggested that a shallow Convolutional Neural Network obtained the best performance in both conditions, outperforming state-of-the-art methods on the used data sets. This work was published as a work-in-progress in the CHI conference .
BCPy, an open-source python platform for offline EEG signals decoding and analysis: Although promising, BCIs are still barely used outside laboratories due to their poor robustness. Moreover, they are sensitive to noise, outliers and the non-stationarity of EEG signals. Many algorithms have been developed for EEG signals processing and classification, in order to improve BCIs robustness. We proposed BCPy, an open-source, easy-to-use python BCI platform for offline EEG signal analysis. Python is free and contains good scalable libraries for scientific computing. Moreover, Python is the major language used to implement recent advances in ML and Deep Learning, thus making them easily available for BCI research. This work was published in the International BCI meeting .
Adaptive Riemannian classification methods: The omnipresence of non-stationarity and noise in EEG signals restricts the ubiquitous use of BCIs. One of the possible ways to tackle this problem is to adapt the computational model used to detect and classify different mental states. Adapting the model will possibly help us to track the changes and thus reducing the effect of non-stationarities. In this paper, we present different adaptation strategies for state of the art Riemannian geometry based classifiers. The offline evaluation of our proposed methods on two different datasets showed a statistically significant improvement over baseline non-adaptive classifiers. Moreover, we also demonstrate that combining different (hybrid) adaptation strategies generally increased the performance over individual adaptation schemes. Also, the improvement in average classification accuracy for a 3-class mental imagery BCI with hybrid adaption is as high as around 17% above the baseline non-adaptive classifier. This was published in .
Regularized spatial filters for EEG regression problems: In collaboration with University Freiburg, we reported on novel supervised algorithms for single-trial brain state decoding. When brain activity is assessed by multichannel recordings, spatial filters computed by the source power comodulation (SPoC) algorithm allow identifying oscillatory subspaces. In small dataset scenarios, this supervised method tends to overfit to its training data. To improve upon this, we proposed and characterize three types of regularization techniques for SPoC. Evaluating all methods on real-world data, we observed an improved regression performance mainly for datasets from subjects with initially poor performance. This was published in the Neuroinformatics journal .
SEREEGA: a toolbox to Simulate EEG activity: EEG is a popular method to monitor brain activity, but it is difficult to evaluate EEG-based analysis methods because no ground-truth brain activity is available for comparison. Therefore, to test and evaluate such methods, in collaboration with TU Berlin, we proposed SEREEGA, a free and open-source matlab toolbox for Simulating Event-Related EEG Activity. The toolbox is available at https://github.com/lrkrol/SEREEGA. SEREEGA unifies the majority of past simulation methods reported in the literature into one toolbox. This toolbox and its use were published in journal of neuroscience methods .