Authors:

Assadollahi, R. & PULVERMULLER, F.

Reference:

Year of publication:

2001

CBU number:

5181

Abstract:

Can a self-organising neuronal network classify neuromagnetic recordings following presentation of single words, despite the low signal-to-noise-ratio of the MEG? There is evidence that semantic word properties are mirrored in spatio-temporal characteristics of the (averaged) brain response they elicit (Pulvermüller 1999). In addition, amplitude and/or latency of components in the physiological brain response have been shown to reflect word length and frequency (Osterhout et al. 1997). It is possible that differential magnetoencephalographic response are present even in the signal evoked by single word stimuli. Stimulus words from four categories were chosen: function words, action verbs, nouns primarily eliciting visual associations, and nouns with both strong visual and action associations. Word length (one/two syllables) and frequency (high/low) were also varied in an orthogonal design. The sixteen words (4 categories x 2 lengths x 2 frequencies) were presented visually to a single subject whose brain response was recorded by an 148 channel MEG. A neuronal net (Kohonen 1982) was trained on the field strength of 8 regions of interest covering a large part of the cortical surface. After learning, the network performed well above chance on new testing data from the same individual. In the recognition of individual words from the neuromagnetic signal its recognition rate was 25% above chance (c2 = 16.3, p<0.0001) and its accuracy was 30% above chance (c2 = 40.8, p<0.0001). The classification of brain responses into word categories was also unexpectedly high (recognition rate 18% above chance, c2 = 27.2, p<0.0001, accuracy 20% above chance, c2 = 42.0, p<0.0001). We conclude that unsupervised Kohonen maps can classify single word-elicited brain responses.