Multivariate MEG Pattern Classifiers for Language Impairment In Autism

Background: Language impairment (LI) is an important behavioral component of Autism Spectrum Disorders (ASD). Two measures of auditory processing revealed by magnetoencephalography (MEG) that may relate to LI in ASD are M100 and mismatch field (MMF) latencies. Roberts et al. (2010) demonstrated that the latency of the M100 in the superior temporal gyrus was increased in children with autism and could be used to distinguish between autistic and typically developing (TD) subjects. This measurement, however, could not distinguish between language-impaired (LI+) and non-language-impaired (LI-) subjects. The MMF latency, a measure of how quickly the brain detects changes among sounds or phonemes, is lengthened in ASD and may also be a measure of LI (Roberts et al., 2008). While these univariate measures characterize some aspect of LI/ASD, we hypothesize that their appropriate combination will improve the group distinction.

Objectives: Using multivariate machine-learning pattern classifiers created by combining the two MEG latency measures, we evaluate whether this combination can distinguish between ASD/LI+, ASD/LI- and TD populations better than univariate classifiers, by learning patterns of disease variability, incorporating age effects and generating an abnormality score that can aid diagnosis of an unseen subject.

Methods: We measured MMF and M100 latencies in the superior temporal gyrus in 41 children with ASD (13 LI+, 28 LI-, defined according to clinical/neuropsychological assessment) and 21 TD controls using MEG. In our data (as in several other studies: Gage et al., 2003b; Oram Cardy et al., 2008) M100 correlated significantly with age (R=-.5129, p=2.0x10^-5), with no significant difference in this correlation between groups. Hence, M100 measurements were age-corrected by linear regression. Using MMF and age-corrected M100 as input features, we created a non-linear two-way support vector machine (SVM) classifier of ASD and TD populations and a three-way classifier between the ASD/LI+, ASD/LI- and TD groups, cross-validated with a leave-one-out paradigm. For each classifier, we obtained an average classification accuracy by testing on each left out subject using the constructed classifier. For the three-way classifiers, the sensitivity for LI+ was recorded as the percentage of LI+ individuals that were correctly classified.

Results: The two-way ASD-TD classifier using MMF and M100 yielded a cross-validated classification accuracy of 83.87% (52/62) with sensitivity of 87.8% (36/41) and specificity of 71.4% (15/21) compared to univariate MMF classifier (accuracy 80.65% (50/62); sensitivity 90.24% (37/41); specificity 61.9% (13/21)) and M100 classifier (accuracy 75.81% (47/62); sensitivity of 95.12% (39/41); specificity 38.1% (8/21)). The three-way multivariate classifier yielded a cross-validated accuracy of 66.13% (41/62), significantly higher than the chance rate of 33% (p<0.001). The sensitivity towards LI+ was 38.46% (5/13). The classifier using only M100 yielded an overall accuracy of 59.68% (37/62), but sensitivity towards LI+ was inferior  (7.69%, 1/13; p=0.06), with similar values for a 3-way univariate MMF classifier.

Conclusions: This work presents multivariate MEG-based classifiers that provide better group separation between ASD and TD populations compared to univariate M100 and MMF classifiers. Further, the three-way multivariate classifiers (despite small sample sizes) were able to better elucidate the heterogeneity of ASD, distinguishing degrees of LI.