Auditory imagery classification with a non-invasive Brain Computer Interface

Brain Computer Interfaces (BCIs) controlled by motor imagery have been explored during many years mainly because of its ease of imagination, and localization of cortex activity. Auditory imagery, nonetheless, has been a rarely explored approach, but a one that might open new possibilities in musical interfaces, communication and speech synthesis. It can also bring a complementary way to motor imagery for improving accuracy in BCIs. Using a non-invasive BCI, open source software, and auditory imagery of white noise, we tested several classification algorithms to determine the accuracy and usefulness of using non-motor imagery EEG signals. We tested 15 healthy adults with a 6 electrode EEG setup and the open source platform OpenVibe. The signals were analyzed with MultiLayer Perceptron, Linear Discriminant Analysis, and Support Vector Machines classifiers. Our results show that using a Support Vector Machine classifier and our specific experimental setup, we could achieve up to 93% accuracy in white noise imagery in our subjects (Linear Discriminant Analysis and MultiLayer Perceptron setups yielded accuracy of 73-76%). Thus, auditory imagery can be a promising complementary approach to traditional motor-imagery BCIs.