Discipline: Technology and Engineering
Subcategory: Biomedical Engineering
Joseph K. Nuamah - North Carolina Agricultural and Technical State University
Co-Author(s): Younho Seong, North Carolina Agricultural and Technical State University, Greensboro, NC
The application of autonomous systems is on an increase, and there is the need to optimize the fit between humans and these systems. While human operators must be aware of the autonomous system’s dynamic behaviors, the autonomous system must in turn base its operations, among other things, on an on-going knowledge of the human operator’s cognitive state, and its application domain. Human roles will shift significantly to a supervisory one in human autonomous systems interactions. The ability to unobtrusively monitor levels of task engagement could be useful in identifying more accurate methods for humans to interact with autonomous systems. Although previous studies are suggestive of the potential of classifiers using physiological features to determine the level of cognitive activity in tasks, some questions still remain. In particular, can Task Engagement Indices (TEIs) be used as features for classification of cognitive tasks? We address this issue by employing TEIs as input to Support Vector Machines (SVM) and Artificial Neural Networks (ANN) to allow identification and classification of mental engagement, and compare their classification accuracies. We assert that differences in task demand will elicit different degrees of mental engagement, which could be measured through the use of the TEI. Electroencephalography (EEG) was acquired from six healthy participants with an electro-cap elastic electrode cap during five cognitive tasks. At a sampling rate of 250Hz, data was recorded for 10 seconds during each task and each task was repeated five times per session. Data from six participants were used. All data analyses were performed offline. The short term Fourier transform was used to estimate the power spectrum of the EEG bands. For each participant under each of the five cognitive tasks, for each trial, and for each of the six channels, the TEI was computed. This resulted in a total of six features per trial per task. For the ANN classification, six separate feedforward ANN with single hidden layer trained by backpropagation were designed to classify the five mental tasks for each participant. The performance function used for all the six ANN networks was the mean square errors. The best performing single layer feedforward ANN for each participant was determined by varying the number of hidden nodes from 10 to 50 in steps of 5. For the SVM classification, six separate SVM with radial bias function kernel were designed to classify the five tasks. Across all six participants, the ANN and SVM achieved an average overall classification accuracy of 88.67% and 93.33% respectively. These results indicate that EEG task engagement index may serve as features for identification and classification of mental engagement of cognitive tasks. Future research involves exploring other representation techniques of EEG signals, such as discrete wavelet transform, in order to reliably discriminate and understand extracted relationships.
Funder Acknowledgement(s): This work was partially supported by DoD Army Research Office (W911NF-13-1-018), and by Air Force Research Laboratory (FA8750-15-2-0116) through Younho Seong.
Faculty Advisor: Younho Seong, yseong@ncat.edu
Role: I did all the research. I wrote the importance of the research, did the methods and controls, wrote the results, and arrived at the conclusions and future research.