Toward estimation of brain connectivity as new feature for BCI application

Introduction Modern neuroimaging has provided unequivocal evidence that brain functions are subserved by multiple areas functionally interconnected. Brain computer interfaces (BCIs) may benefit from feature extraction based on metrics derived from brain connectivity [1]-[3]. However, reaching the minimally required accuracy when few data samples are available as in single trial or real-time connectivity applications is still challenging. Variables selection algorithms could represent a valuable alternative to the classical algorithms, since they provide accurate connectivity estimates even when the amount of data samples available is very scarce [4]. The aim of the present work is to propose an accurate and reliable approach for connectivity estimation that paves the way to the use of features extracted from a connectivity analysis for BCI applications. Methods and Results A. Variable selection approaches When few data samples are available for multivariate analysis, it is necessary to use estimators based on variable selection approaches. In this works, we selected LASSO, Group LASSO, Fused LASSO and Elastic Net algorithms [5]. The idea at the basis of variable selection techniques is to select only those connectivity parameters that have an effect on the response vector. The remaining parameters are set to zero. B. Simulation study Algorithms were tested on simulated data (N=50) according to the following steps: i. Generation of simulated EEG datasets, fitting predefined ground truth networks of 10 nodes under different condition of factor DL (80, 120, 160, 240, 320, 480), where DL is the number of data samples available for the estimation process. ii. Estimation of connectivity parameters by means of variable selection methods (factor TYPE: LASSO, E-NET, F-LASSO, G-LASSO). iii. Evaluation of their performances by means of MAPE (for parameters estimation) and AUC parameters (for the validation) [6], [7]. MAPE represents the error committed in estimating the connection strength, while AUC measures the accuracy in the assessment of null and non-null connections. A repeated measures ANOVA was computed on MAPE and AUC parameters considering as within factors DL and TYPE. C. Results As reported in Figure 1, we found that the MAPE parameter (panel a) decreased with the increasing of the number of data samples available (F(15,735)=47.1, p<10-4). Instead, the AUC parameter (panel b) showed the opposite trend (F(15,735)=106.2, p<10-4). Even if the trends were observed with all the algorithms tested, LASSO showed the best performances (lowest MAPE and highest AUC for all amount of data samples considered). Discussion The presented findings obtained by applying variable selection approaches demonstrated the possibility to reliably estimate brain connectivity based on few data samples. Despite the boundary conditions in which all the algorithms were analysed, they tend to the optimal values of MAPE (0%) and AUC (1) with the increase of number of data samples available for the estimation process. Noticeably, LASSO regression showed the best performances for both estimation and validation procedures as compared with other considered algorithms. Even if variable selection approaches have already been used for brain connectivity