Using Machine Learning to Improve Motor Imagery Neurofeedback

Abstract

Machine Learning (ML) was employed to identify features in magnetoencephalography (MEG) recordings of 16 participants performing motor imagery (MI). ML was applied to data obtained using three methods: 1) two sources localized to somatosensory cortex; 2) all 306 MEG sensors; and 3) 80 anatomical sources localized via known coordinates. A linear kernel support vector machine was fit to frequency features extracted by fast fourier transform and accuracy (ability to accurately class trials as rest or imagery) evaluated with leave one batch out cross validation. The two-source model performed poorly with accuracy of ~0.7, while the 80-source and sensor-level models performed well with accuracy greater than 0.9. Sensor-level models selected sensors analogous to C4, FC4, and F10 as key sensors. Use of ML to identify features in MI-based neuroimaging data can be applied to make MI interventions clinically feasible options through reducing cost, setup time, and expertise required to employ neurofeedback.