Support Vector Machine (SVM) Analysis of Auditory Oddball Event-Related Potentials (ERP) Classifies Toddlers with and without Early Signs of Autism

Background:  

ERPs have been used to characterize sensory and cognitive function in autism (Jeste & Nelson, 2009; Marco et al, 2011). In particular, differences between children with and without autism have been noted in responses to auditory (speech) oddball paradigms. Recently, attempts have been made to identify biomarkers for autism risk via the sophisticated analysis of complexity in resting state electroencephalogram (EEG) signals of young children with and without risk factors for autism (Bosl et al, 2011). Preliminary findings suggest that the use of multiclass SVM of EEG data is a promising approach to the classification of infants with risk factors for neurodevelopmental disorders. We propose that application of similar analysis techniques on auditory oddball ERP data where known differences exist between autism and non-autism, should be more successful in classifying groups and isolating specific biomarkers.

Objectives:  

The purpose of our study was to examine the utility of SVM analysis of auditory oddball ERP data in the classification of toddlers (12-24 months) with and without early signs of autism.

Methods:  

Forty-six toddlers (mean age=17.9 months, SD=3.0, 28 males) participated in the study. ERPs were collected using an EGI GES 300 system utilizing a HydroCel 128 Channel Geodesic Sensory Net and Net Amps 300 amplifier. Toddlers completed an auditory oddball paradigm involving phonemes (dae and daa; stimulus duration = 340ms, ISI = 960ms). Stimuli were presented in 4 x blocks of 400 stimuli each lasting approximately 8.5 minutes. Data was processed in Net Station and included: (1) filtering (high-pass=0.1 Hz and low-pass=30Hz), (2) segmentation, (3) artifact detection, (4) bad channel replacement, (5) referencing and (6) baseline correction.

Participants were assessed as showing early signs or no early signs of autism using the Autism Detection in Early Childhood (ADEC; Young, 2007) screening tool. Toddlers scoring between 0-5 were identified as no early signs (N-ES) (n=24) and toddlers scoring a 6 and above were identified as showing early signs (ES) (n=22).

Results:  ERP data from 20 ES toddlers and 18 N-ES toddlers was submitted for preliminary analysis to SVM with Gaussian and linear kernels for classification purposes. Feature vectors representative of each child were generated by averaging the raw time series data from a single channel over all standard and deviant responses in a block. The performance of the classifier was analyzed using leave-one-out cross-validation.

Conclusions:  

Preliminary results from our study indicate that SVM analysis of auditory oddball ERP data correctly classified three-quarters of our sample of toddlers with and without early signs of autism. Further analysis of the full dataset using data from multiple channels is expected to strengthen this result. Analysis techniques such as these may be key in isolating subtle differences in sensory and cognitive development associated with autism.