Classifying children with reading difficulties from non-impaired readers via symbolic dynamics and complexity analysis of MEG resting-state data

Magnetoencephalography (MEG) is a brain imaging method affording real-time temporal, and adequate spatial resolution to reveal aberrant neurophysiological function associated with dyslexia. In this study we analyzed sensor-level resting-state neuromagnetic recordings from 25 reading-disabled children and 27 non-impaired readers under the notion of symbolic dynamics and complexity analysis. We compared two techniques for estimating the complexity of MEG time-series in each of 8 frequency bands based on symbolic dynamics: (a) Lempel-Ziv complexity (LZC) entailing binarization of each MEG time series using the mean amplitude as a threshold, and (b) An approach based on the neural-gas algorithm (NG) which has been used by our group in the context of various symbolization schemes. The NG approach transforms each MEG time series to more than two symbols by learning the reconstructed manifold of each time series with a small error. Using this algorithm we computed a complexity index (CI) based on the distribution of words up to a predetermined length. The relative performance of the two complexity indexes was assessed using a classification procedure based on k-NN and Support Vector Machines. Results revealed the capacity of CI to discriminate impaired from non-impaired readers with 80% accuracy. Corresponding performance of LZC values did not exceed 55%. These findings indicate that symbolization of MEG recordings with an appropriate neuroinformatic approach, such as the proposed CI metric, may be of value in understanding the neural dynamics of dyslexia.