| |
Abstract:
We investigate the prediction of a motor response, prior to
motor action, through analysis of MEG signals. Subjects performed a
visual-motor integration task, learning to press a button with
their left or right hand based on a visual stimulus. Data was
recorded for 4 subjects across 90 trials. MEG signals were
collected using a 122 sensor NeuroMag system. Subjects learned the
task after one or two trials. Sampling was done at 300Hz. MEG
signals were analyzed in a 70msec window, 30msec before the button
press. Predictions were made with a logistic regression (LR) model,
trained and tested using leave-one out. Input to the LR model was a
122 element vector, each element being the magnitude of one of the
MEG sensors in a 3msec interval. LR model was trained to predict
left button push. Performance of the LR model was analyzed using
receiver operating characteristic (ROC) analysis. Average area
under the ROC curve for the 4 subjects was Az=0.85 (std=0.06).
Localization of the LR discrimination vector, using an inverse
dipole fitting algorithm, showed localization to the contra-lateral
hemisphere in motor/sensory cortex. When mapped to the
motor-sensory homunculus, the LR discrimination vector localized to
the hand/thumb region. Our results indicate prediction of motor
commands is possible through analysis of MEG and that learned
discrimination vectors are consistent with the cortical functional
architecture.
|
