Abstract:
Based on the current Diagnostic and Statistical Manual (DSM 5) (American Psychiatric Association, 2013), Autism spectrum disorder (ASD) is a behaviourally categorized syndrome, generally determined early in childhood, consisting of mild to severe social and communication deficits, perseverative and rigid ideation, often accompanied by various stereotypies (Whyatt, 2017). ASD is a life-long condition without any effective intervention to improve the core deficits (Juane Heflin & Simpson, 1998). Diagnosis of ASD is difficult as the diagnosis relies upon behavioural signs and symptoms and, to date, no biological distinctions identifiy ASD (Loth et al., 2016). Despite technological advances, currently approved medical treatment for ASD consists of behavioural training, drug-based symptom management and environmental accommodations (Cardon, 2016). Depending on the severity, people with ASD often find ways to cope and adapt to their deficits. The majority of people who are mildly impaired are able to live relatively normal lives, but the rest do not at great cost to both them and their communities. Over the past few decades, recognition of a need to improve the diagnosis and treatment of ASD has led to novel methods of evaluation, including the use of electroencephalography (EEG) applying various algorithms that employ fractal dimension (FD), entropy and other complex analytic methods. These more recent approaches to EEG analysis are facilitated by innovative mathematics that commenced over five decades ago as well as recent increases in calculation capacity of computers. We use the term ‘complexity analysis’ here to encompass fractal, entropy and other non-linear measures based on the model of the brain as a complex dynamic system (CDS) which depart from commonly used linear analysis. The purpose of this thesis is to employ complexity analysis of EEG signal data to attempt to distinguish higher risk ASD adults from lower risk controls and determine if a distinction can be made that is consistent with behavioural measures, related to the diagnostic criteria. Previous studies have reported that complexity measures can differentiate between resting states, tasks and conditions such as ASD better than linear measures. To avoid confounding our results by preprocessing the EEG data, as is common to remove artefacts and line noise that can interfere with the analysis, we ran both raw and two versions of preprocessed EEG as a control. Our research comprises three studies. Study 1 is a pilot investigation of EEG complexity features with three adult participants to distinguish a hierarchy between three conditions of a resting state (RS): eyes open (EO), eyes closed (EC) and during the transition between (EOEC), to compare six complexity measure algorithms: the correlation dimension (CD) that establishes a peak fractal embedding dimension, calculates the peak dimension and checks it with the false nearest neighbours method (CDPKFNN), the largest Lyapunov exponent (LLE), Higuchi’s fractal dimension (HFD), multi-scale entropy (MSE), multi-fractal detrended fluctuation analysis (MFDFA), and Kolmogorov complexity using Lempel-Ziv methodology (KC). Study 2 is a pilot study expanding on the first investigation to distinguish a hierarchy between three task conditions: a resting state—EO; a cognitive task—the Berg card sorting test (BCST) used for neuropsychological evaluation; and a social skills task—Reading the Mind in the Eyes Test (RMET) used to screen for ASD risk. These three task conditions were compared using the same six complexity measure values. Based on complexity values, Study 3 employs all six measures to distinguish between 39 adult participants who were scored on the autism-spectrum quotient (AQ), to compare a higher or lower risk for ASD. We also investigate some of the possible causes of inconsistent results in our studies and those found in the literature. For Study 1, we tested our hypothesis that the EO state was more complex and values would be higher in all measures by applying the Wilcoxon Signed Rank Test (WSRT) to compare probabilities between median same channel values. Our results for all three pilot participants showed CDPKFNN values were highest in the EO state in raw i.e. un-preprocessed EEG, but for HFD, EO values were lowest in preprocessed EEG. We found that no values for other measures were highest or lowest for any resting state and trends of higher or lower measures in two participants for one or more tests were mixed. This showed that, similar to some mixed results in the literature, EO raw EEG appeared to be a more complex as measured by CDPKFNN, while preprocessed EEG without low-pass band filtering had the lowest value when measured by HFD. In contradiction to the literature we found the other measures had no lowest or highest RS values. We attribute this result to a potential issue with HFD parameters that may not have been optimized for this data or a complexity feature difference between measures. We speculate that other measures failed to show a statistically significant difference between resting states due to a lack of parameter optimization for the data set and the small number of participants. In Study 2 we tested our hypothesis that EEG during the RMET was more complex and would be higher in all measures applying the WSRT to compare probabilities between median same channel values. We found several measures showed task differences with all three participants. CDPKFNN values showed that BCST had the lowest values in raw and preprocessed EEG without low-pass band filtering, with EO trending lower for raw EEG. Interestingly, LLE results for all tasks for all three participants showed a hierarchy of RMET>EO>BCST for preprocessed EEG, with BCST also lowest for all preprocessing setups and no trends. However, as in Study 1, HFD values went in the opposite direction, with RMET having the lowest values in raw EEG. Yet, MSE values, like LLE, set up the same hierarchy of RMET>EO>BCST for preprocessed EEG with trends for lower EO without low-pass band filtering. MFDFA had no highest values for any task but both BCST preprocessed and RMET without low-pass band filtering were lowest, with mixed trends. Finally, KC had RMET lowest for all three participants in the raw EEG setup with BCST trending higher. The overall results were mixed yet appeared to trend in favor of RMET as highest for both LLE and MSE; BCST lowest in CDPKFNN, LLE, and MSE; but with HFD and KC having RMET as lowest. We again speculate differences in complexity features, a lack of parameter optimisations and the small sample size contributed to these mixed results. For Study 3 we tested our hypotheses that overall or relative values would be lower for the higher AQ score group versus lower score group. First, we examined the frequency distributions for our behavioural measures and found some significant correlations between BCST perseverative scores and AQ scores, as well as a small correlation between RMET and AQ scores. For the EEG data, using WSRT to compare probabilities between median same channel values, we found no correlations or trends between AQ score and any of the measures in either of the two groups. Based on inspection of the boxplots and subsequent review of the pvalues, no pattern emerged that indicated a hierarchy of tasks overall based on AQ score. This was an interesting result given some indications from the pilot studies that a hierarchy may exist and the reporting in the literature of higher values for certain cognitive tasks. Due to the small number of participants in the pilot studies, the result is not inexplicable as there was a possibility of apparent significant hierarchy by chance (~ 3%). Despite its common practice in the literature, we did not select channels and frequencies that favored our hypothesis to avoid a type I error. We consider this early research in a relatively new area of EEG analysis that suggests further exploration, as many values were close to the significance threshold and optimization of parameters may improve results.