Spatial Correlation Preserving EEG Dimensionality Reduction Using Machine Learning

• Published in: 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) pp. 2583 - 2589
• Main Authors: Gebre-Amlak, Haymanot, Nguyen, Hoang Mark, Lowe, Jesse, Nabulsi, Ala-Addin, Chuy, Narisa Nan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2018
• Subjects: Brain modeling; Correlation; Electrodes; Electroencephalography; Support vector machines; Task analysis; Training
• DOI: 10.1109/BIBM.2018.8621106

An electroencephalogram (EEG) is a recording of the brains electrical activity measured at the scalp. It is composed of multifaceted signals that reflect the localized activity and intensity of neurological processes opening the possibilities for brain to machine interaction. This research explores the use of Machine Learning to analyze EEG data collected from a custom-designed NeuroRacer virtual interactive cognitive training software. The method utilizes Machine Learning classification based on channel level feature performance to significantly reduce complexity while maintaining tolerable loss in performance prediction. The EEG data were collected using a Model"Active-Two" sensor cap (from Cortech Solutions) in conjunction with a BioSemiActiveTwo 64-channel EEG acquisition system. Our findings indicate that classification of the signal feature, derived from the full array of 64 channels during a 6-month follow-up, yielded a prediction rate of 82.2% when assessing the intervention induced differentiation between a Multi-Task Training (MTT) and Non-contact Control groups. Furthermore, we reveal that using a smaller number of channels where the success rate of prediction is optimized would lead to practical benefit for the EEG device design. Our results show that we reach the optimal level with channel reduction down to a selective 12 out of 64 channels with a penalty of 10.04% error rate.