Deep learning approach to semi-automatic feature selection in BCI using expert-annotated labels

Feature selection is a fundamental step in the implementation of the BCI closed loop architecture. One of the features selection methods currently used involves the manual inspection and selection of features starting from Fisher scores feature maps computed from the PSD of the EEG signal. One major problem with this approach is that it is time consuming, and feature selection can be biased by the operator, making the process subjective. The goal of this work is to develop and test a possible neural network model that can perform a semi-automatic feature selection starting from the same data available to experts when they manually perform the feature selection process. To this end, for each of the available EEG signals in the dataset, 192 PSD frequency channel pair features have been computed. Through crowdsourcing, answers on the features to select for each sample have been collected and used to generate soft labels that incorporate uncertainty in the feature selection process that emerged from the responses received. The results obtained on the test set indicate the potential for a model that enables semiautomatic feature selection and contributes to reducing subjectivity bias. Nevertheless, the clear errors committed by the model show that there is still a margin of improvement and the model cannot be used on its own but it still needs some sort of supervision of an expert. In addition, to verify the impact of the features selected by the model when there are discrepancies with respect to the ground truth, LDA models were developed and tested to simulate the performances in a pseudo-online loop in order to have a comparison between the results when using experts-selected features and model-selected features. The results of the LDA classifiers using the features selected by experts versus those proposed by the neural network model—and restricting the analysis to cases where the model did not exhibit clear feature selection errors—indicate that the neural network was capable of identifying alternative features not chosen by experts, which could potentially enhance the performance of the LDA classifiers. Nevertheless, this finding should be interpreted with caution and considered only as a preliminary hypothesis, since the observed discrepancies between the model and the experts may also arise from model inaccuracies that coincidentally led to the selection of features with apparent interpretative relevance.