Evaluation of classification performance of functional near infrared spectroscopy signals during movement execution for developing a brain–computer interface application using optimal channels

• Published in: Journal of near infrared spectroscopy (United Kingdom) 2018-08, Vol.26 (4), p.209-221
• Main Authors: Janani, A, Sasikala, M
• Format: Article
• Language: English
• ISSN: 0967-0335; 1751-6552
• DOI: 10.1177/0967033518787331

The objective of brain–computer interface application is to give a new lease of life to motor-impaired people by enabling them to communicate with the external world purely based on thoughts. Functional near infrared spectroscopy technology uses light to detect brain activations related to the motor activity performed by the subjects and an emerging modality in the brain–computer interface field. The study aims at comparing haemodynamic responses during upper limb movement tasks and resting condition over 14 channels placed equally on right and left hemispheres. Channel selection based on statistical methodology has been introduced to optimally select those channels that give statistical differences between the motor tasks. Besides using commonly used features such as mean, variance and peak of oxygenated haemoglobin signal, two other features, namely difference between mean concentration changes of oxygenated haemoglobin and deoxygenated haemoglobin and difference between mean activation level of oxygenated haemoglobin during the task condition and mean activation level of oxygenated haemoglobin during the resting state, are investigated. Their effect on the artificial neural network classifier accuracy over the selected channels is analysed. The proposed channel selection strategy is compared with the channel-averaging approach, which is one of the conventional channel selection methods. Also, performance of these methods is compared when all 14 channels are considered for classification. Classification of functional near infrared spectroscopy signals corresponding to motor tasks and rest condition gave accuracy values of (87.25 ± 8.66)% over 14 channels, (83.81 ± 5.97)% over channels selected using channel-averaging method and (84.71 ± 4.97)% over the selected subset of four channels. Hence, the results in terms of classification accuracies are promising and demonstrate the prospect of developing functional near infrared spectroscopy-based brain–computer interface application. Also, the analysis carried out in the study showed that the upper limb motor tasks evoked different haemodynamic responses in the reduced set of channels located over centro-parietal regions of the brain.