AutoBCI: interface cérebro-máquina com configuração hiperparamétrica automatizada

Abstract

Motor Imagery-based Brain-Computer Interfaces (MI-BCI) allow control of devices without the use of peripheral nerves and muscles, based on voluntary modulation of brain electrophysiological activity. The challenge imposed on the typical non-invasive MI-BCI is to extract patterns that describe the motor intention in signals collected by electroencephalography (EEG) and classify them to generate reliable commands to the application. For that, the selection of suitable processing techniques as well as the correct parameterization of the system are fundamental in the adjustment of effective classification models. The configuration of multiple hyperparameters in the processing chain, commonly performed manually and unspecified by the user, tends to generate rigid models that are unable to generalize well in different individuals, especially due to the high variability of MI patterns observed among them. The use of strategies to estimate these hyperparameters according to the subject’s specificities is presented as a more effective approach and has been explored in recent studies. This work proposes a structure based on Bayesian learning incorporated into a new open source MI-BCI computational platform for automatic configuration of hyperparameters. The system integrates all the basic steps of the ICM-IM subband architecture, from the acquisition to the control of a virtual application. Various processing techniques make up a large configuration space to search for particular hyperparametric instances that maximize system performance and draw the user the manual adjustment task. Data from 72 subjects in three public EEG sets were used in offline and online simulations, whose goal was to validate the operation of the implemented modules and to investigate the effects of the automatic configuration on the classification performance and on the effective control of the application. A significant improvement in the accuracy of classification was observed when using automatic configuration based models of the system compared to models generated from frequent configurations in the literature. The results suggest that the optimization of hyperparameters produces more assertive models in the classification of IM patterns of different users and tends to contribute to a more effective control of the application. It is concluded that this study contributes to the design of ICM-IM more effective in recognizing the user’s particular IM patterns by providing a complete experimental environment, customizable and uncomplicated to use by automated configuration. The option for more efficient techniques in signal processing also proved to be viable and are also considered contributions of this work.