Machine-learning-enabled adaptive signal decomposition for a brain-computer interface using EEG

Abstract

Background and objectiveThe use of adaptive signal decomposition methods and machine learning (ML) algorithms have gained interest in biomedical applications. Brain-computer interfaces (BCIs) are one such application that uses electroencephalogram (EEG) signals to assist people with severe brain disabilities to communicate with their external environment. This paper uses recent adaptive signal decomposition methods such as empirical mode decomposition, empirical wavelet transform, variational mode decomposition, and variational non-linear chirp mode decomposition to decompose EEG signals into various modes. Three publicly available EEG datasets are employed. EEG datasets are decomposed into modes using the aforementioned methods. Linear and non-linear time-domain features are extracted from the modes. A feature set is obtained by selecting highly discriminant features using the Kruskal-Wallis test. Classification is performed using four recent ML-based algorithms. ResultsThe maximum classification accuracy achieved is 89.6 ± 4.6% and 61.1 ± 5.1% in binary and multiclass (seven classes) signals, respectively, for dataset 1. The effectiveness of the proposed model is measured by evaluating the performance parameters such as parameters recall (REC), specificity (SPEC), precision (PREC), F1-score, and area under the curve (AUC). For binary classification, REC, SPEC, PREC, F1-score, and AUC are 95%, 94%, 38%, 55%, and 89% (95% CI: [0.871–0.937]), respectively. Also, for multiclass classification, REC, SPEC, PREC, F1-score, and AUC are achieved as 60%, 93%, 60%, 60%, and 89% (95% CI: [0.873–0.971]), respectively. ConclusionsThe results show that the employed algorithms achieve excellent accuracy for the available datasets. The results also strengthen the possibility of using EEG signals for silent communication.