Recurrence Quantification Analysis of Resting State EEG As Risk Biomarker for Non-Syndromal and Syndromal ASD

Background:  There has been growing interest in electroencephalography (EEG) as an investigational tool for biomarker development in autism spectrum disorder and other neurodevelopmental disorders. However, one of the key challenges lies in the identification of appropriate multivariate, next-generation analytical methodologies that can characterise the complex, nonlinear dynamics of neural networks in the brain. Recurrence quantification analysis (RQA) may be a potential next-generation approach for the identification of individuals ‘at risk’ of ASD. A proof of principle study, comprising 7 ASD and 5 typically developing (TD) subjects (8-17 years), showed that linear discriminant analysis (LDA) of RQA features could classify subjects as ASD or TD with 83.3% accuracy, 85.7% sensitivity and 80.0% specificity. The sample size was, however, small, and did not consider age, gender and intellectual ability as potential confounders. Furthermore, a robust biomarker for ASD should be able to differentiate ASD not only from TD, but also from other neurodevelopmental aberrations, as seen in genetic syndromes, for example.  

Objectives:  Here we present further exploration of this method in a larger dataset comprising non-syndromal ASD, Tuberous Sclerosis Complex (TSC) with ASD (syndromal ASD), TSC without a concurrent diagnosis of ASD, and TD subjects.   

Methods:  Raw resting state EEG (rsEEG) data were filtered and preprocessed, artefact contaminated segments were manually rejected and semi-automated ocular artefact correction was performed using independent component analysis. Data were extracted in continuous 5 second segments. RQA features were extracted from recurrence plots of multivariate embedded rsEEG data. LDA, multilayer perceptron (MLP) neural network and support vector machine (SVM) classifiers were used to classify the RQA feature vector. A leave-one-subject-out approach was utilised to simulate the diagnosis of an unseen subject. Biomarker performance was evaluated in three samples, each matched for age, gender and intellectual ability, where possible. Multiple segments per subject were evaluated to determine test-retest reliability. Study 1 analysed 666 segments from 7 ASD (mean age 3.96 years) and 7 TD (mean age 3.93 years) subjects. Study 2 analysed 1202 segments from 5 TSC+ASD (mean age 5.92 years) and 5 TSC-ASD (mean age 5.99 years) subjects. Study 3 analysed 832 segments from 6 ASD (mean age 3.90 years) and 6 TSC+ASD (mean age 3.89 years) subjects.   

Results:  In study 1, 92.9% accuracy, 100% sensitivity, and 85.7% specificity was achieved using an SVM classifier. Study 2 showed 90% accuracy, 80% sensitivity, and 100% specificity, using a MLP classifier. In study 3 both the SVM and MLP classifiers achieved 100% accuracy, sensitivity and specificity. 

Conclusions:  Results suggested that RQA may be a reliable approach to classify individuals with syndromal and non-syndromal ASD. Age, gender and intellectual ability were identified as potential confounders for classification performance in poorly matched samples. With comparison of the non-syndromal and TSC+ASD sample population, epilepsy and medication use were also identified as possible confounding factors. Rigorous investigation of each of these factors in a well matched and larger sample population will be required with further biomarker development. Consistent test-retest reliability will also need to be established.