Study of Algorithms for Implementation of Brain-Computer Interfaces in Embedded Systems

The implementation of a Brain-Computer Interface (BCI) system requires electroencephalography (EEG) signal processing that includes filtering, feature extraction, and classification algorithms. As such, the present work proposes to use the PocketBeagle embedded system to run algorithms for BCI using the Python language. This work aimed to explore some features of the embedded system to optimize performance and resource consumption, as well as the training time of the implemented algorithms, which used Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) classifiers, both with Common Spatial Patterns (CSP) filter. When comparing previous research results with the developed algorithms of this work embedded in the PocketBeagle, the training time increased by 42.98 s for LDA and 42.66 s for SVM. When analyzing the memory consumption of the implementations in the embedded system, the codes consumed less than half of the memory available in the 512 MB PocketBeagle. The consumption of the LDA classifier was 167 MB at its peak, and the SVM was 177 MB at the peak of its execution. Using the metrics resulting from the confusion matrix, it is clear that the SVM classifier had a better performance than the LDA since its accuracy is 83.14 % and its f-score is 0.8111, while for the LDA classifier, they are 66.29 % and 0.6940, respectively.