What Do We Currently Know about Resting State EEG Biomarkers in Autism Spectrum Disorder?

Background:  Electroencephalography is a noninvasive technique that captures the underlying electrical changes in brain activity at the scalp surface. Interdisciplinary research endeavours are beginning to show promise in the development of resting state electroencephalography (rsEEG) biomarkers for the early detection of ASD. However, limited information is available about the similarities and differences between analysis methods used to date, and about the gaps in knowledge about these methods.

Objectives:  Here we conducted a critical review of the current state of rsEEG biomarkers in ASD to draw attention to clinical and methodological limitations that need to be addressed in future work.

Methods:  A systematic review was conducted using “resting state”, “EEG”, “biomarker”, and “autism”, contained in the title, keywords or abstracts of articles in BioMed Central, PubMed, Scopus, ScienceDirect and IEEE Xplore journals. Primary papers identified were used to identify secondary literature sources regarding strengths and weaknesses of identified methods.

Results:  Three primary methodological papers were identified: modified multiscale entropy (MME) by Bosl et al. (2011), coherence analysis (CA) by Duffy and Als (2012), and recurrence quantification analysis (RQA) by Pistorius et al. (2013). Bosl et al. (2011) implemented a nonlinear complexity univariate feature extraction technique able to distinguish infants at high risk for ASD and typically developing controls; a follow-up investigation incorporating the final diagnoses of each child is pending. It is anticipated that features extracted using a multivariate feature extraction technique, i.e. from all EEG channels combined, will provide information relating to the system as a whole and will enable the extraction of more informative features that would enable better group discrimination. Duffy and Als (2012) implemented univariate feature extraction of coherence features to distinguish children with and without ASD; this approach employs Fourier analysis which assumes that data are stationary and that analysis of short segments is sufficient. Long EEG segments, typically minutes in length, are required in order to obtain reliable coherence estimates. A trade-off thus exists between segment length and stationarity – the segment length must be long enough to yield good frequency resolution, but short enough to satisfy the assumption of stationarity. The RQA feature extraction methodology proposed by Pistorius et al. (2013) was the third promising method for biomarker development, as RQA can be applied in univariate or multivariate time series analysis, it can reliably analyse short segments, and can be applied to linear or nonlinear data without having to make prior assumptions regarding linearity or stationarity of data. The RQA method requires evaluation on larger-scale samples.

Conclusions:  Scrutiny of the three biomarker methods reported to date suggest that, whilst binary categorical classification of ASD versus typically developing children may be possible, many other questions remain unanswered. No studies to date have examined biomarkers for ASD in relation to other neurodevelopmental disorders, for instance. It is likely that rsEEG biomarkers will be sensitive to a range of factors that require rigorous evaluation, including age, gender, eyes-open versus eyes-closed condition, or the presence of artefacts.