Autism, spectrum or clusters? An EEG coherence study

The DSM-5 [1] summarizes that individuals on the autism spectrum exhibit problems involving interaction and communication with other individuals, show repetitive behaviors and restricted interests, and manifest behavior issues interfering with school, work, and/or multiple other life endeavors. The move from DSM-3 [2] to DSM-4 [3] and most recently DSM-5 [1] as diagnostic standard reflects a gradual condensation of a number of autism-related clinical entities under the rubric of Autism Spectrum Disorder (ASD). These include infantile autism, atypical autism, pervasive developmental disorder not otherwise specified or PDD-nos, and most recently Asperger’s syndrome [4]. This diagnostic “simplification” was welcomed by some yet quite concerning to others, as previously reviewed [5]. As suggested by Kienle et al. [6], “…the issue of an empirically reproducible and clinically feasible differentiation into subgroups must still be raised.” Indeed, in 2016, Pruett and Povinelli [7] published a wistful ‘Commentary’ in which they hypothesized that the usual rapid and automatic recognition of individuals on the autism spectrum resulted from our human “evolved sensitivity for species-typical ranges of social relating”. The authors further postulated that “social spacing”, “quality of eye contact”, and “timing of communicative exchange” constitute three primary variables and that they may form a recognizable set of “clusters” within the realm of human behavior.

A review of the literature from 1994 to 2018 reveals nine publications using cluster analysis to demonstrate quantitatively defined groupings within subjects diagnosed with autism spectrum disorder (ASD) [6, 8–15]. Eight papers used structured neurobehavioral assessments of various types [2, 3, 16–24] and one relied upon MRI data. Four studies reported solutions involving two clusters, one reported both two and three cluster solutions, one study reported a three cluster solution, two reported a four cluster solution, and one reported a five cluster solution.

It is notable, that all studies summarized above succeeded in identifying clusters. However, the number of underlying clusters identified varied, although the two cluster solution was noted most often. As an ensemble, the studies serve to suggest that ASD may well comprise a varying number of discrete sub-populations rather than exist on a continuum.

The varying number of clusters reported in these different studies may reflect the unique characteristics of the population under study, differing choices of variables selected to represent subjects and/or differing cluster methodologies utilized. The two most commonly used cluster methods, hierarchical and K-means algorithms rely upon apparent ‘satisfactory’ cluster separation by means of the clinical/neuropsychological difference between or among subjects within differing clusters. K-means clustering, hierarchical clustering, and combinations of these two fundamental methods, fail to determine quantitatively the optimal cluster number. For instance, the K-means approach to clustering “…requires users to specify the number of clusters to be generated. One fundamental question is: How to choose the right number of clusters” [25], p 39. Similarly, “…one of the problems with hierarchical clustering is that is does not tell us how many clusters there are…” [25], p 74.

The current study employed one of the first comprehensive software approaches to objectively establish cluster number, namely NbClust [25]. It is specifically designed to provide an objective means, i.e. independent of investigator choice, to identify the ‘optimal’ cluster number within a population.

Thus, the main goal of the current study was to determine the feasibility of delineating objective, EEG-based, clusters among children diagnosed with ASD. The underlying hope is that successful clustering might ultimately lead to better clinical, cost effective, specific diagnoses of subtypes of Autism and the creation of more specific interventions as well as a method to test the effectiveness of such interventions, be they pharmacological, behavioral or other. The multiplicity of potential approaches to clustering and their complexities have been succinctly reviewed by Jain et al. [26] and more recently by Charrad et al. [25]. Key issues that arise when employing cluster analysis are as follows: (1) What clustering technique to use (typical choices are K-means and hierarchical); (2) How many clusters to form (typically required prior to analysis initiation); and (3) How to determine the relative “significance” of resulting cluster configurations (internally by statistics and/or externally by association with one or more symptom complexes). The software package NbClust [25] addresses these three issues and was used in this study to elucidate the cluster structure within a large group of ASD subjects each represented by 40 previously derived [27] EEG-based coherence factors.

The main hypothesis of the study thus was that there are definable subgroups (clusters) of children within autism. It was hypothesized that children in one cluster will be different from children in other groups (clusters), and more similar among each other within a cluster, than across clusters. A secondary hypothesis states that EEG coherence is a productive means for the establishment of such stable clusters of children within the autism population.

Results show that 430 subjects diagnosed as being on the autism “spectrum” and represented by 40 EEG coherence factors [27], fell into two distinct clusters. These two ‘autism clusters’ statistically differed from one another and, in turn, statistically differed from 554 subject neuro-typical control group subjects, not involved in the clustering process. Notably the 40 utilized EEG coherence factor variables had been objectively derived [27] and a completely objective data-driven variable selection was applied. Furthermore, choice of the optimal cluster number was also objectively determined by use of a relatively recent software package, NbClust [25]. This program was instructed to form up to 15 clusters and to establish the optimal cluster configuration on the basis of the 30 methods [25] included in the program.

Finally, NbClust was run twice, first utilizing the hierarchical clustering technique and second utilizing the other commonly used K-means technique. Both techniques ‘voted’ the two cluster configuration as optimal; the choice was more definitive when hierarchical clustering was used. Thus, the optimal two cluster solution was selected on the basis of objectively derived EEG measures of brain connectivity.

In order to explore the potential clinical significance of the two autism clusters, advantage was taken of a prior study [5] that contrasted the control, ASD, and ASP populations and that had shown that ASP subjects were closer to the ASD population than the neurotypical control population, and also that ASP subjects were statistically separable and distinctly different from the ASD population. For the current study, these previously studied 26 ASP subjects were represented by the 40 EEG coherence factor variables and were utilized to determine whether these ASP subjects would passively fall within one or the other of the newly formed clusters. Notably these ASP subjects had not been utilized in the original cluster formation. As Fig. 7 shows, the majority of ASP subjects were within or close to the Cluster 2 domain, which suggests that C2 may be primarily associated with those subjects manifesting Asperger-like behavioral characteristics. It is of note that despite the multiple variable types and differing methods for clustering in the literature, a prominent two cluster distribution of autistic characteristics has been observed repeatedly by others [6, 8, 11, 14, 15].

A significant limitation of the study is the lack of external validation by similarly extensive subject data from other relevant domains, such as neuro-psychological evaluation as well as autistic specific evaluations such as the ADOS [32] or ADI-R [45], MRI [9, 46–49], genetic/epigenetic testing [50–57], and prolonged sleep EEG recordings for detection of epileptiform activity [29, 58–61].

The future establishment of correlations among such additional brain-based data with the EEG coherence-based findings described in this manuscript should be very helpful in further validating the EEG-based subgroups, and should facilitate interpretation of the coherence data by clinicians and scientists alike. The absence of correlative data in the current study does not invalidate the future use of the current data and findings for such correlative studies. For example it was possible in the current study to insert 26 Asperger’s patient data within the three group discriminant analysis described here. As also previously reported, autistic patients could be additionally classified as also having attention issues using a discriminant, developed on a different population with attention disorders. Thus it was possible to explore attention disability within autism [40] by means of EEG data. Future studies utilizing the current study’s results would only require that subjects have waking EEG data. Such studies are anticipated.

It is also important to clarify that the current demonstration of two neurophysiological clusters within the autism spectrum does not preclude the possibility of further, relevant autism subdivisions. For example in the future, as the neurodevelopmental characteristics of Cluster 1 and 2 are explored, there may prove to be additional clinically relevant sub-populations within or even across these two clusters. .

Objective brain derived EEG coherence factor data strongly support the proposition that the autism disorder should not be seen as a continuous spectrum [1] but likely is formed from at least two distinct subpopulations. This is important since the ‘spectrum versus clusters’ issue goes beyond academic taxonomy and has a number of real world consequences: (1) For example, moving Asperger’s syndrome subjects into the autism spectrum disorder allows some US schools to develop and offer a single autism educational program that is oriented towards management and teaching of the ‘typical’ autistic child of limited verbal ability, who may present with behavioral issues. This may leave out or otherwise disadvantage children with Asperger’s Syndrome, who present with specific and often different behavioral and educational issues altogether and typically profit from more individually-tailored education. (2) On the other hand, in clinical autism research (e.g., neuro-behavioral evaluations, MRI based studies) it is often much easier to recruit and successfully study subjects with Asperger’s Syndrome or others who are high-functioning. However, results of such studies may be inappropriately generalized as findings characteristic of the entire autism spectrum. (3) The multiple different findings resultant from genetic and epigenetic studies of autism [54, 62–65] also contradict a unitary perspective on autism. Important correlational findings may be lost when all autism is treated as a single entity. It prevents identification of distinct subgroups based upon clinical insights, and/or neurobehavioral parameters, and/or direct brain parameters (as from MRI and EEG).

In addition EEG, a classic and relatively inexpensive non-invasive test, is found to be reasonably well tolerated by children with various forms of ASD and should be considered for inclusion in future studies of autism and of other neurobehavioral disorders [5, 27, 40, 66, 67].

As previously discussed [27] the authors believe that in clinical practice diagnosis of ASD should follow the DSM-5 criteria, and should be made by clinicians with special training in this area (e.g., neurologists, psychiatrists, psychologists) by use of readily available assessments such as the ADI-R [45]. EEG coherence study data may best serve as adjunctive, confirmatory, and/or exploratory information. At this point they are especially useful regarding the discovery of clinically relevant autism sub-populations.