Decoding perceptual thresholds from MEG/EEG

Abstract

Abstract-Magnetoencephalography (MEG) can map brain activity by recording the electromagnetic fields generated by the electrical currents in the brain during a perceptual or cognitive task. This technique offers a very high temporal resolution that allows noninvasive brain exploration at a millisecond (ms) time scale. Decoding, a.k.a. brain reading, consists in predicting from neuroimaging data the subject's behavior and/or the parameters of the perceived stimuli. This is facilitated by the use of supervised learning techniques. In this work we consider the problem of decoding a target variable with ordered values. This target reflects the use of a parametric experimental design in which a parameter of the stimulus is continuously modulated during the experiment. The decoding step is performed by a Ridge regression. The evaluation metric, given the ordinal nature of the target is performed by a ranking metric. On a visual paradigm consisting of random dot kinematograms with 7 coherence levels recorded on 36 subjects we show that one can predict the perceptual thresholds of the subjects from the MEG data. Results are obtained in sensor space and for source estimates in relevant regions of interests (MT, pSTS, mSTS, VLPFC).

Data-driven HRF estimation for encoding and decoding models

“…We argue that this evaluation metric is better suited than a regression loss for this task because of the discrete and ordered nature of the labels. Also, this loss is less sensible to shrinkage of the prediction that might occur when penalizing a regression model (Bekhti et al, 2014). The Kendall tau coefficient always lies within the interval [−1, 1], with 1 being perfect agreement between the two rankings and −1 perfect disagreement.…”