Circuit Dynamics and Computational Neuroscience I.1.l Deep and machine learning Monday AM + Wednesday AM

2473 - Multi-class Word Imagery Speech BCI Classification by Machine Learning, Operational Architectonics and Complex Networks

Topic / Sub Topic
I.1.l Deep and machine learning

Abstract

Abstract Body

Aims: Speech imagery of electroencephalographic (EEG)-based Brain Computer Interface (BCI) is significant for people with motor disabilities, illnesses and speech disorders. However, a reliable and efficient performance of these BCI systems depends strongly on the classification accuracy of speech imagery. Therefore, the development of more robust and consistent classification methods is needed for improving communicating imagined speech BCI systems.

Methods: Imagery pronunciation of three words (“in”, “out”, “up”) was performed by five subjects for 100 trials. EEG data sets of 64 electrodes were recorded. The EEG data sets, which are publicly available [1], were analyzed based on a novel classification algorithm comprised of “three pillars”: a) Operational Architectonics concept of brain and mind functioning [2]. b) Complex network measures of brain connectivity. c) Machine Learning for developing multi-class classifiers.

Results: The results show that the overall mean classification accuracies ranged approximately from 53% to 66.5%, significantly above chance level (33.33%) in all tested cases.

Conclusions: Even though classifying imagined speech from EEG data is a difficult task, this study demonstrates that is feasible to recognize the features which distinguish words, such as “in”, “out” and “up”, from information embedded in the EEG signals, with fairly good accuracies. Further analysis of new EEG-based BCI data sets is needed to verify the aforementioned results.

References

1. Nguyen et al. (2018) J. Neural Eng. 15:016002 (16pp).

2. Fingelkurts and Fingelkurts (2015) Neuromethods 91: 1–59.

Hide