Automatic identification of received language in MEG

We identify the language being received during English and Romanian auditory stimuli in 11 subjects before and after a period of learning 50 words in the latter using only magnetoencephalographic measures. To accomplish this, we extract on the order of 100,000 features (based on wavelets and descriptive statistics over windowed signals), and identify the most salient features. While we achieve very high accuracy in pre-training (up to 90% mean accuracy across 10-fold cross-validation for some subjects), it is significantly more difficult to tell received languages apart after training. We also identify significant effects of semantic word category and the subject’s ability to play a musical instrument on classification accuracy.