Probability Distributions.- Linear Models for Regression.- Linear Models for Classification.- Neural Networks.- Kernel Methods.- Sparse Kernel Machines.- Graphical Models.- Mixture Models and EM.- Approximate Inference.- Sampling Methods.- Continuous Latent Variables.- Sequential Data.- Combining Models.

Pattern Recognition and Machine Learning

Co-Planar Stereotaxic Atlas of the Human Brain—3-Dimensional Proportional System: An Approach to Cerebral Imaging, J. Talairach, P. Tournoux. Georg Thieme Verlag, New York (1988), 122 pp., 130 figs. DM 268

The power of feedback.

An overview of the self-organizing map algorithm, on which the papers in this issue are based, is presented in this article.

The Self-Organizing Map

In this paper we review classification algorithms used to design brain–computer interface (BCI) systems based on electroencephalography (EEG). We briefly present the commonly employed algorithms and describe their critical properties. Based on the literature, we compare them in terms of performance and provide guidelines to choose the suitable classification algorithm(s) for a specific BCI.

/pdf/a-review-of-classification-algorithms-for-eeg-based-brain-p1h22hb1wr.pdf

A review of classification algorithms for EEG-based brain–computer interfaces

Mental-Imagery based Brain-Computer Interfaces (MI-BCIs) make use of brain signals produced during mental imagery tasks to control a computerized system. The current unreliability of MI-BCIs could be due, at least in part, to the use of inappropriate user-training procedures. In order to improve these procedures, it is necessary first to understand the mechanisms underlying MI-BCI user-training, notably through the identification of the factors influencing it. Thus, we aim at creating a statistical model that could explain/predict the mean performances of MI-BCI users using their traits (e.g., personality) but also the evolution of performances using demographic data (age, gender) and the timing of the experiment (time lapse between two sessions). We used the data of 42 participants collected from three different studies [1-3] that were based on the same MI-BCI paradigm. They were asked to learn to control an MI-BCI by performing three MI-tasks (i.e., left-hand motor imagery, mental rotation and mental subtraction) across different training sessions (3 or 6 depending on the experiment). We used a LASSO regression (Least Absolute Shrinkage and Selection Operator)[4] with a leave-one-subject-out cross validation to build different models. Our first results showed that using the users' traits may only enable the prediction of performances within one multiple-session experiment, but might not be sufficient to reliably predict MI-BCI performances across experiments. In a second time, we were able to find a model gathering all the subjects that could predict the mean performance of a session using the participant's gender, the timing of the experiment and the mean performance over the previous session (p < .01). Further studies considering, for instance, an estimation of the users' states and new metrics to assess performances are necessary to reveal more reliable models of MI-BCI performances.

https://hal.archives-ouvertes.fr/hal-02432459/document

Computational modelling to predict/explain MI-BCI users' performances and their progression

OBJECTIVE
The usage of Riemannian geometry for Brain-computer interfaces (BCIs) has gained momentum in re-cent years. Most of the machine learning techniques proposed for Riemannian BCIs consider the data distribution on a man-ifold to be unimodal. However, the distribution is likely to be multimodal rather than unimodal since high-data variability is a crucial limitation of electroencephalography (EEG). In this paper, we propose a novel data modeling method for considering complex data distributions on a Riemannian manifold of EEG covariance matrices, aiming to improve BCI reliability.


METHODS
Our method, Riemannian spectral clustering (RiSC), represents EEG covariance matrix distribution on a manifold using a graph with proposed sim-ilarity measurement based on geodesic distances, then clusters the graph nodes through spectral clustering. This allows flexibility to model both a unimodal and a multimodal distribution on a manifold. RiSC can be used as a basis to design an outlier detector named outlier detection Riemannian spectral clustering (oden-RiSC) and a multimodal classifier named multimodal classifier Riemannian spectral clustering (mcRiSC). All required parameters of odenRiSC/mcRiSC are selected in data-driven manner. More-over, there is no need to pre-set a threshold for outlier detection and the number of modes for multimodal classification.


RESULTS
The experimental evaluation revealed odenRiSC can detect EEG outliers more accurately than existing methods and mcRiSC out-performed the standard unimodal classifier, especially on high-variability datasets.


CONCLUSION
odenRiSC/mcRiSC are anticipated to contribute to making real-life BCIs outside labs and neuroer-gonomics applications more robust.


SIGNIFICANCE
RiSC can work as a robust EEG outlier detector and multimodal classifier.

Modeling complex EEG data distribution on the Riemannian manifold toward outlier detection and multimodal classification.

Stroke is the leading cause of complex disability in adults. The prevalence of motor deficit and cognitive impairment after stroke is high and persistent. The most common consequence is the hemiparesis of the contralateral upper limb, with over 80% of stroke patients suffering from this condition acutely and over 40% chronically. Brain-Computer Interfaces (BCI) based on motor imagery have shown promising results in post-stroke motor recovery. However, this approach does not work for all patients, and even when it works, shows vastly different effectiveness across patients. It thus needs to be improved. This could be achieved by personalizing the BCI-based Motor Rehabilitation (MR) program to each patient, notably by personalizing the employed Artificial Intelligence (AI) models used. To do this, it is necessary to first identify the predictive factors of successful BCI-based motor rehabilitation. In fact, very little research has addressed the question of factors that influence post-stroke BCI-based MR. Thus, in this paper, we present a survey of the literature about the factors related to successful use of BCIs in general and then the factors that are associated to post-stroke motor recovery, to identify the various factors that could influence BCI-based post-stroke MR. We then discuss how such factors could be taken into account in order to develop new AI algorithms for personalized post-stroke BCI-based MR.

Identifying factors influencing the outcome of BCI-based post stroke motor rehabilitation towards its personalization with Artificial Intelligence

When using Brain-Computer Interfaces (BCI) based on ElectroEncephaloGraphy (EEG), the identification of mental tasksrelies on two main points: feature extraction and classification [MAM+06, BFWB07, LCL+07]. Feature extraction aims atdescribing EEG signals by a few relevant values called “features ”, whereas classification aims at automaticallyassigning a class to these features. In this paper we focus on feature extraction, as the BCI community has stressed theneed to explore new feature extraction algorithms [MAM+06].Recently, inverse models have been revealed as promising feature extraction algorithms for BCI [QDH04, GGP+05,WGWB05, CLL06]. Such models aims at computing the activity in the whole brain volume, by using only scalp EEGsignals and a head model representing the brain as a set of voxels (volume elements). The activity computed in a fewbrain regions has been used as features for BCI systems.Despite good results, some limitations remain. Indeed, it seems that current methods cannot conciliate genericity, i.e., thecapability to deal with any kind of mental task, and the fact of generating few features. On one hand, methods that aregeneric and automatic tend to generate a large number of features, as they extract several features for each voxel[GGP+05]. The activity in neighboring voxels can be correlated and, as such, it would be more appropriate to gather thesevoxels in brain regions. On the other hand, methods that generate few features have been proposed, but they are notgeneric anymore as they need a priori knowledge on the mental tasks used, and are currently limited to motor imagerybasedBCI [QDH04, WGWB05]. Recently, we have proposed a method which is generic and which generates few features,as voxels whose activity is correlated are gathered into Regions Of Interest (ROI) [CLL06]. However, this method is notcompletely automatic and is limited to the use of two ROI whose spatial extension is hard to define [CLL06].In this paper, we propose a generic feature extraction algorithm which can automatically identify any number of relevantROI and can properly define their spatial extension thanks to the new concept of fuzzy ROI. This algorithm is known asFuRIA, which stands for “Fuzzy Region of Interest Activity”.The FuRIA algorithmFuRIA is a trainable feature extraction algorithm. Indeed, it can automatically identify what are the relevant ROI todiscriminate mental tasks, and what are the frequency bands in which these ROI react. The FuRIA training is performed in 5 steps:Statistical analysis : This step consists in comparing, using a statistical test, the activity in each voxel, possiblyaveraged over a given time window, between the different classes. This analysis is performed for each of the Nffrequencies considered. Once the analysis is achieved, we obtain Nf statistical maps, each one revealing which voxelsare significant for discrimination.Clustering : For each statistical map computed at the previous step, a clustering algorithm is used in order to gathersignificant voxels into ROI. To perform this clustering, a 4-dimensional vector [x,y,z,s] is associated to each significantvoxel. The x,y,z coordinates are the 3D coordinates of the voxel in the head model used and s is the voxel statisticobtained at the previous step. Once the clustering is achieved, all voxels whose associated vector belongs to the samecluster are gathered into the same ROI.Fusion : Once the relevant ROI are identified for each frequency, we have to identify which ROI are relevant in whichfrequency band. To do so, ROI obtained at similar positions (i.e., with an overlap of more than 50 %), for consecutivefrequencies, are merged, simply using the union set operator, and identified as reactive in the frequency band resultingfrom the concatenation of these frequencies.Adaptation : Because of the use of the union as the fusion operator, a ROI identified as reactive in the frequencyband [ fa; fb] can contain voxels which were significant for a single frequency belonging to [ fa; fb] but which are notsignificant anymore for the whole [ fa; fb] frequency band. Consequently, we perform another statistical analysis foreach ROI, in order to remove voxels that are not significant anymore.Fuzzification : In the ROI defined so far, it can be noticed that all voxels do not have, statistically, the samediscriminative power. However, all these voxels contain a given amount of information. Moreover, the brain regions of agiven subject are not well defined but are rather naturally fuzzy. Hence, identifying which voxels should be kept in a ROI,and which ones should be removed is a difficult problem. We believe that all voxels that contain information should be inthe ROI, but that those which contain less information should be less in the ROI. Thus, we propose the concept of fuzzyROI, which relies on the fuzzy set theory of Zadeh [Zad96]. A standard (crisp) ROI c is defined by the set of voxels itcontains. The activity γc in this ROI c is classically computed from the activity γv of the voxels v of this ROI (seeEquation 1). A fuzzy ROI f is not defined by a set of voxels anymore, but by a fuzzy membership function μf. Thisfunction provides the membership degree (in [0; 1]) of any voxel into the ROI f. The activity in a fuzzy ROI is computedas described in Equation 2, where Nv is the number of voxels in the head model used. This formalism enables use toweigh each voxel according to the importance of its contribution within the ROI, and, as such, to use efficiently the availableinformation. To complete the training of FuRIA, all ROI are turned into fuzzy ROI by associating to each of them afuzzy membership function μ which uses the statistics obtained at the previous step.Once the training of FuRIA is completed, a set of discriminant fuzzy ROI should have been identified together with thecorresponding reactive frequency bands. The activity in these ROI and in these frequency bands, possibly averaged overtime, can be used as features for any classifier.traitement du signal 2008_volume 25_numero 4 Colloque GRETSI 2007 307FuRIA : un nouvel algorithme d’extraction de caracteristiques pour les interfaces cerveau-ordinateur utilisant modeles inverses et modeles flousEvaluationIn order to evaluate FuRIA, we used sLORETA (standardized low resolution electromagnetic tomography) as the inversemodel [PM02], a multiple comparison randomization t-test as the statistical analysis [HBWF96] and the “Mean Shift”algorithm as the clustering algorithm [CM02]. Concerning the fuzzy membership functions, it appears that, due to thelow resolution of sLORETA, significant voxels with a relatively high probability of type I error are much more numerousthat significant voxels with a relatively low probability of type I error. Consequently, for each ROI , we used amembership function μ which assigns a membership degree which decreases exponentially with the voxel statistics,as described in Equation 10. In this equation, Ns is the number of voxels in the ROI ; pmin and pmax are respectivelythe minimum and maximum probability, among the ROI voxels, of correctly rejecting the null hypothesis H0 ; and pv isthe probability of correctly rejecting H0 for voxel v.In order to evaluate FuRIA, we have used the data set IV of the “BCI competition 2003 ” [WZL+04]. These data containEEG signals recorded 500 ms before a subject hits a keyboard key with a finger from his left or right hand. The goal ofthe competition participants was to identify, for each trial, the hand that was used.We trained the FuRIA algorithm on the training set provided, by considering the frequencies between 3 Hz and 45 Hz,with a step of 1 Hz, and by considering the time window comprising the last 250 ms of each trial. FuRIA has found 5discriminant fuzzy ROI (see Figure 1), hence generating 5-dimensional feature vectors.These features has been classified using a linear support vector machine. The recognition rates obtained after 100training/testing repetitions are displayed in Table 2 and compared with two other methods : the method of thecompetition winner [WZL+04] and a method using inverse models that we proposed previously [CLL06]. The resultsobtained withouth the fuzzification step (step 5 of the training, see previous section) are also displayed as a comparison.We can noticed that the method that use FuRIA with fuzzification reached the best results with a recognition scoreranging from 85 % to 86%. These results suggest that the features extracted with FuRIA are very discriminant, as theyare not numerous (there are only 5 features) and still enable us to reach good classification results. It is also interestingto notice that when using FuRIA without fuzzification, the results are significantly lower. This stress the advantage ofusing efficiently the information contained in any significant voxel, thanks to the concept of fuzzy ROI.In addition to this evaluation, we also proposed a first step towards the design of an interpretable BCI system. To thisend, we proposed to classify FuRIA features, which are interpretable features, with a Fuzzy Inference System (FIS), whichis an interpretable classifier. Indeed, FIS are a set of fuzzy “if-then” rules that can be learnt from data. By combiningFuRIA features with FIS, we can obtain a set of rules describing which activity in which brain regions and frequencybands is related to which mental task. An evaluation of this method on the EEG data mentioned above suggested thatthe obtained rules were consistent with the literature (see Figure 3).

FuRIA : un nouvel algorithme d’extraction de caractéristiques pour les interfaces cerveau-ordinateur utilisant modèles inverses

Systeme d'affichage de donnees caracterisant l'activite cerebrale d'un individu, comportant : - un bloc de capture cerebrale adapte pour collecter des donnees caracterisant l'activite cerebrale dudit individu a chacun d'une pluralite de pas de temps ; ledit systeme d'affichage etant caracterise en ce qu'il comprend un bloc afficheur affichant une image actualisee dudit individu a chaque pas de temps ; et en ce qu'il est adapte pour detecter, a chaque pas de temps, une zone, correspondant a une section predefinie de l'individu, de ladite image affichee de l'individu et pour commander l'affichage de l'image d'un objet superposee sur ladite zone de l'image affichee de l'individu, certaines caracteristiques dudit objet affiche etant fonction des donnees collectees au dit pas de temps.

Fabien Lotte

Papers

Computational modelling to predict/explain MI-BCI users' performances and their progression

Modeling complex EEG data distribution on the Riemannian manifold toward outlier detection and multimodal classification.

Identifying factors influencing the outcome of BCI-based post stroke motor rehabilitation towards its personalization with Artificial Intelligence

FuRIA : un nouvel algorithme d’extraction de caractéristiques pour les interfaces cerveau-ordinateur utilisant modèles inverses

Systeme d'affichage de donnees caracterisant l'activite cerebrale d'un individu, procede et programme d'ordinateur associes