Content uploaded by Kristina Aldridge
Author content
All content in this area was uploaded by Kristina Aldridge on Aug 06, 2015
Content may be subject to copyright.
ORIGINAL PAPER
Facial Structure Analysis Separates Autism Spectrum Disorders
into Meaningful Clinical Subgroups
Tayo Obafemi-Ajayi •Judith H. Miles •T. Nicole Takahashi •Wenchuan Qi •
Kristina Aldridge •Minqi Zhang •Shi-Qing Xin •Ying He •Ye Duan
ÓSpringer Science+Business Media New York 2014
Abstract Varied cluster analysis were applied to facial
surface measurements from 62 prepubertal boys with
essential autism to determine whether facial morphology
constitutes viable biomarker for delineation of discrete
Autism Spectrum Disorders (ASD) subgroups. Earlier
study indicated utility of facial morphology for autism
subgrouping (Aldridge et al. in Mol Autism 2(1):15, 2011).
Geodesic distances between standardized facial landmarks
were measured from three-dimensional stereo-photogram-
metric images. Subjects were evaluated for autism-related
symptoms, neurologic, cognitive, familial, and phenotypic
variants. The most compact cluster is clinically character-
ized by severe ASD, significant cognitive impairment and
language regression. This verifies utility of facially-based
ASD subtypes and validates Aldridge et al.’s severe ASD
subgroup, notwithstanding different techniques. It suggests
that language regression may define a unique ASD sub-
group with potential etiologic differences.
Keywords Autism Cluster analysis Language
regression Facial phenotype Biomarker Outcome
indicators
Introduction
Autism Spectrum Disorder (ASD) comprises a group of
complex neuropsychiatric disorders of childhood, diag-
nosed on the basis of the behavioral phenotype. The ASD
phenotype is characterized by social deficits, impaired
Some preliminary result from this work is being presented as a poster
presentation at the 2014 International Meeting for Autism Research
(IMFAR) in Atlanta, Georgia, USA on May 14, 2014.
Present Address:
T. Obafemi-Ajayi
Applied Computational Intelligence Lab, Department of
Electrical and Computer Engineering, Missouri University of
Science and Technology, 301 W. 16th St, Rolla, MO 65409,
USA
T. Obafemi-Ajayi W. Qi Y. Duan (&)
Department of Computer Science, University of Missouri, 201
Engineering Building West, Columbia, MO 65211, USA
e-mail: duanye@missouri.edu
J. H. Miles T. N. Takahashi
Thompson Center for Autism and Neurodevelopmental
Disorders, University of Missouri, 205 Portland Street,
Columbia, MO 65211, USA
J. H. Miles
Department of Child Health, University of Missouri School of
Medicine, One Hospital Dr, N712, Columbia, MO 65212, USA
K. Aldridge
Department of Pathology and Anatomical Sciences, University
of Missouri School of Medicine, One Hospital Dr, M309 Med
Sci Bldg, Columbia, MO 65212, USA
M. Zhang S.-Q. Xin Y. He
School of Computer Engineering, Nanyang Technological
University, 50 Nanyang Avenue, Singapore 639798, Singapore
Present Address:
S.-Q. Xin
College of Information Science and Engineering, Ningbo
University, 818 Fenghua Road, Ningbo, Zhejiang, China
123
J Autism Dev Disord
DOI 10.1007/s10803-014-2290-8
communication, and restricted and repetitive behavior
patterns (American Psychiatric Association 2013). A recent
study by Aldridge et al. (2011) using 3D facial imaging
discovered structural differences between faces of children
with ASD and typically developing children. They sug-
gested that differences in facial morphology may reflect
alterations in embryologic brain development. Within ASD
they identified two clinically discrete ASD subgroups,
using cluster analysis of the facial measurements. This
current study validates the previous findings by using an
alternative distance measurement and multiple clustering
techniques to verify the power and utility of facially based
ASD subtypes. Geodesic (surface) rather than Euclidean
(straight) measurements and four exceptionally robust
clustering techniques were utilized to determine whether
similar or additional subgroups would be identified.
Extensive use of mathematical algorithms for data selec-
tion, multiple cluster analysis techniques, validity and
classification models optimized the results.
Experimental results from cluster analysis based on
facial morphology using surface distance features revealed
that the ASD cohort studied could be separated into three
clusters. Examination of clinical data using mean and
correlation analysis revealed that each of the three clusters
demonstrated relatively distinctive clinical and behavioral
traits. One of the clusters (Cluster 2) exhibited clinical
traits similar to those described by Aldridge et al. (2011)in
their subgroup 1. If these facial groups identify etiologi-
cally discrete subsets of ASD, their identification may
allow clinicians and researchers to identify precise etio-
logic bases of the ASD. This study demonstrates the gen-
eralization of facial phenotypes as a viable biomarker for
identifying ASD subgroups. The similarity of the results
obtained show that it is not dependent on measurement
type (Euclidean vs. geodesic) or the cluster technique. This
confirms that facial measurements provide a replicable and
important biomarker in autism.
Methods
Subjects
Sixty-two prepubertal Caucasian boys between 8 and
12 years of age, who had been diagnosed with ASD at the
Thompson Center for Autism and Neurodevelopmental
Disorders, were recruited for study. Forty-two subjects had
also participated in the Simons Simplex Collection (SSC)
and their clinical data were available for analysis. The
remaining 22 subjects were recruited from the Thompson
Center database, which contains similar clinical data. To
ensure a homogeneous study set, all subjects were male, of
Caucasian ancestry and old enough to have a mature facial
and skull growth, but prepubertal to avert androgen surge
effects on facial bone growth (Farkas and Posnick 1992)
and classified as having essential autism. Boys with rec-
ognized genetic syndromes, including fragile X syndrome,
chromosomal disorders, including copy number variants
(CNV), generalized dysmorphology or gestational age less
than 35 weeks were excluded. Generalized dysmorphology
was assessed using the autism dysmorphology measure
(ADM) (Miles et al. 2008). In addition, 83 % (52/63) of the
ASD subjects overlap with the Aldridge cohort (Aldridge
et al. 2011). A control group of 36 typically developing
prepubertal Caucasian boys between 8 and 12 years of age
were recruited from the community under the Thompson
Center control subject recruitment protocol.
ASD Diagnosis
ASD diagnoses were made using the Thompson Center
diagnostic protocol, which consists of complete clinical,
medical, behavioral, and family histories, physical, neuro-
logic and dysmorphology examinations, and autism diag-
nostic measures. Of the 62 boys with ASD, 42 had also
completed the SSC protocol, which included the Autism
Diagnostic Interview—Revised (ADI-R) (Lord et al. 1994)
and Autism Diagnostic Observation Schedule (ADOS)
(Lord et al. 2000). The 20 boys diagnosed exclusively
through the Autism Medical Clinic were diagnosed on the
basis of DSM-IV (American Psychiatric Association 2000)
criteria (as appropriate during the time period of diagnosis)
using a center-specific protocol based on the ADI-R, clin-
ical observation and judgment of the clinician. Seventy-five
percent also had an ADI-R or ADOS, which substantiated
the Thompson Center diagnosis. ASD DSM-IV subtype
diagnoses present within the study population were Autistic
Disorder, Asperger Syndrome, and Pervasive Development
Disorder-not otherwise specified (PDD-NOS).
This study was carried out under the guidelines and
approval of the Health Sciences Institutional Review
Board. The parents or legal guardians of all subjects
Fig. 1 Illustration of the 19 Farkas anthropometric landmark points
used to derive facial surface distance features
J Autism Dev Disord
123
provided written consent for participation in this study;
each subject provided voluntary assent.
Data Acquisition
The 3DMD
Ò
Cranial system was used to reconstruct the
3D surface model (both the geometry and the co-registered
texture image) of each subject, similar to previous work
(Aldridge et al. 2011). We used the 3dMD software to
obtain 3D coordinate data for a set of 19 anthropometric
facial landmarks, as shown in Fig. 1, following (Farkas
1994). These landmark measurements were carried out by a
rater (WQ) trained in use of the software program and
verified by another rater (TO). Facial surface (geodesic)
distances between all possible pairs of the 19 landmark
coordinate points were computed to obtain a total of 171
facial distance features, as described in Fig. 1. For exam-
ple, the distance from the midpoint between both eyes
(B) to the midpoint of the chin (S) is designated as BS
distance feature. Each subject’s facial distance measure-
ments were normalized by dividing them by the geometric
mean of all the geodesic distances obtained for the subject.
3D Geodesic Distance Computation
Geodesic distance is defined as the shortest distance
between any pair of anatomical landmark points along the
surface of the face. Computing geodesics on polyhedral
surfaces has been a fundamental problem in digital
geometry processing and has been extensively studied.
Representative work includes the Mitchell, Mount and
Papadimitriou algorithm (MMP) (Mitchell et al. 1987) and
the Chen and Han algorithm (CH) (Han 1990), which both
compute the exact geodesic distance on triangle meshes. It
is well recognized that this geodesic suffers from topo-
logical and geometric changes due to its local nature. For
example, a small shortcut or miss-measurement may result
in a significantly large change of the geodesic path and
distance. In this work, we apply a global approach for the
robust computation of geodesics on polygonal meshes
(Quynh et al. 2012). This method takes a completely dif-
ferent strategy to compute the geodesic in an iterative and
global manner, in contrast to the MMP and CH algorithms,
which propagate the window (a data structure which
encodes the distance) from the source to the destination.
To compute the shortest distance along the surface, the
first iteration is initialized using the Euclidean distance,
which is able to bridge small holes and gaps. For each
iteration, our method computes the vector field Xwhich
matches the gradient of the current distance field, and
normalizes X. Then it finds the closest scalar potential dby
minimizing RMrdX
jj
2dA over the entire mesh M, which
is equivalent to solve a Poisson equation Dd¼div XðÞ.
These procedures are repeated until the convergence. This
algorithm for Defect-Tolerant Geodesic (DTG) distance
works quite well for the 3D face model, as the computed
geodesics are very resilient to small topological and geo-
metric noises (Xin et al. 2012). Hence, no pre-processing is
required for smoothing or noise removal.
Clinical Data Evaluation
Each of the boys was evaluated for characteristics of their
ASD diagnosis (social function, verbal function, repetitive
behavior and language level), behavioral problems
(aggression, attention deficits and self-injurious behaviors),
out-come measures (IQ, communication, daily living skills,
socialization and Vineland Adaptive Behavior Scale com-
posite scores), the clinical course of their disorder (presence
of regression at onset), medical and neurological variables
(seizures, electroencephalogram results) and physical mor-
phology (head circumference and dysmorphology).
Measures administered include the ADI-R (Lord et al.
1994), ADOS (Lord et al. 1989), Vineland Adaptive
Behavior Scale II (Sparrow et al. 1984), an age- and
development-appropriate IQ test (Full Scale IQ (FSIQ),
Verbal IQ (VIQ), Nonverbal IQ (NVIQ)), Social Respon-
siveness Scale (SRS) (Constantino and Gruber 2005), and
Broad Autism Phenotype (BAPQ) (Sasson et al. 2013).
Parental alcohol use data was obtained using a Parent
Substance Use questionnaire, based on the CAGE
Assessment (Ewing 1984), for families who participated in
the SSC project. A similarly detailed questionnaire was
completed by parents of subjects recruited the Autism
Medical Clinic. Alcoholism was defined as excessive use
of alcohol, tolerance to high amounts of alcohol con-
sumption and/or negative consequences to family, jobs or
health (Wade et al. 2014). In addition, a detailed family
history was obtained for all subjects by the clinician with
extensive experience in the family history method. Not all
measures of IQ were available for a small number of boys.
All participants received complete medical and neurolog-
ical examinations, including assessment of growth and
dysmorphology.
The ADI-R and ADOS, which are considered the gold
standard diagnostic instruments, measure the amount of
impairment for the three autism core symptom areas; 1.
Social functioning, 2. Communication, both verbal and
nonverbal and 3. Repetitive behaviors. Higher scores
indicate greater symptom severity. For each area of
impairment a numeric score specifies the cut-off above
which an ASD diagnosis is indicated. SRS score, devel-
oped for children between 4 and 18 years, measures the
J Autism Dev Disord
123
severity of autism spectrum symptoms that occur in social
settings. It assesses social awareness, social information
processing, capacity for reciprocal social communication,
social anxiety/avoidance, and autistic preoccupations and
traits. BAPQ scores are used to assess relatives of the study
subject for language and personality characteristics diag-
nostic of a broad autism phenotype. The three subscales
quantitatively measure characteristics that correspond to
the diagnosis of autism in the DSM—IV: social deficits,
stereotyped-repetitive behaviors, and social language
deficits.
Cluster Analysis
Cluster analysis is the identification of groups of observa-
tions that are cohesive and separated from other groups
(Fraley and Raftery 2000). Our goal was to identify clusters
of boys with similar facial morphological features within
the ASD dataset that correlate with clinical and behavioral
traits. Our hypothesis is derived from previous work
(Aldridge et al. 2011) that suggested differences in facial
morphology reflect alterations in embryologic brain
development in children with ASD compared to typically
developing children as well as suggesting potential etio-
logic differences. A variety of clustering algorithms can be
used to separate a finite unlabeled data set, like ours, into a
finite and discrete set of ‘‘natural,’’ hidden data structures
(Xu and Wunsch 2005). We chose 4 different clustering
algorithms to apply to our dataset: expectation maximiza-
tion (EM) (Fraley and Raftery 2000), self-organizing fea-
ture map (SOM) (Kohonen 1998), K-means (Hartigan and
Wong 1979), and partitioning around medoids (PAM)
(Kaufman and Rousseeuw 1990).
EM algorithm is a well-known general-purpose machine
learning technique for clustering. It is a model-based
method. EM assigns a probability distribution to each
instance, which indicates the probability of it belonging to
each of the clusters. EM can decide how many clusters to
create by cross validation (as done in all our experiments),
or you may specify a priori how many clusters to generate.
We implemented the EM algorithm using the Weka data-
mining tool (Witten and Frank 2005). SOM also is a
model-based clustering method and uses a neural network
approach. It maps all the instances (points) of a given
dataset in a high-dimensional source space into a 2 to 3-d
target space, such that, the distance and proximity rela-
tionship among the examples in the dataset are preserved as
much as possible. The objective of SOM is to represent
high-dimensional input patterns with prototype vectors that
can be visualized in a two-dimensional lattice structure (Xu
and Wunsch 2005). Each unit in the lattice is called a
neuron, and adjacent neurons are connected to each other,
which provide clear topology of how the network fits itself
to the input space. Input patterns are fully connected to all
neurons via adaptable weights, and during the training
process, neighboring input patterns are projected into the
lattice, corresponding to adjacent neurons. The size of the
lattice, i.e. the number of clusters (k), must be predefined.
K-means is a very simple and widely used partition based
clustering method (Jain 2010). K-means algorithm finds a
partition such that the squared error between the empirical
mean of a cluster and the points in the cluster is minimized.
The goal of K-means is to minimize the sum of the squared
error over all K clusters. K-means algorithm requires three
user-specified parameters: number of clusters K, cluster
initialization, and distance metric. The PAM clustering
algorithm is also a partition based clustering method. PAM
tries to avoid outlier sensitivity, a known fault of K-means,
by using medoids (the most centrally located object in the
cluster) as a reference point rather than the mean value of
the objects in a cluster. Thus, PAM starts from an initial set
of medoids and iteratively replaces one of the medoids by
one of the non-medoids, if it improves the total distance of
the resulting clustering.
Different cluster configurations results were obtained by
varying the input parameters of the four clustering algo-
rithms when applied to the ASD data. However the ques-
tion remains: how do we know which set of clusters is valid
or best fit the data set and how many clusters actually do
exist in the data? Cluster validity refers to formal proce-
dures that evaluate the results of cluster analysis in a
quantitative and objective fashion (Jain 2010). In cluster
validation (Kova
´cs et al. 2005), two measurement criteria
have been proposed for evaluating and selecting an optimal
clustering scheme: Compactness and Separateness. Com-
pactness measures how close the members of each cluster
are to each other. A typical measure of compactness is the
variance. Separateness measures how separated the clusters
are from each other. A good cluster algorithm result should
yield clusters that are compact and well separated. The aim
of cluster validation is to find the cluster partition set which
is the most appropriate/optimal to the input dataset.
The cluster validity analysis platform (CVAP) Matlab
tool (Wang et al. 2009) estimates the quality of the dif-
ferent clustering algorithms’ results and attempts to deter-
mine statistically which set of clusters are optimal using
multiple validity indices. There are different types of
cluster validity indices that measure the quality of clus-
tering results. Validation indices based on internal criteria
assess the fit between the structure imposed by the clus-
tering algorithm (clustering) and the data by itself. Thus,
the clustering results are evaluated using the quantities and
features inherent in the data set (Arbelaitz et al. 2013). For
the evaluation of the multiple clustering results obtained on
the ASD study population, we used four following internal
criteria validation indices to measure the goodness of the
J Autism Dev Disord
123
clusters, since the underlying structure of the data is
unknown.
1. Silhouette index (Rousseeuw 1987). This is a com-
posite index that measures both the compactness (using
the distance between all the points in the same cluster)
and separation of clusters (based on the nearest
neighbor distance). A larger average Silhouette index
indicates a better overall quality of the clustering
result.
2. Dunn index (Halkidi et al. 2001). A measure that
maximizes the inter-cluster distances while minimizing
the intra-cluster distances. A large value indicates the
presence of compact and well-separated clusters. Thus,
the maximum value is the optimal clustering result.
3. Davies-Bouldin (DB) index (Bolshakova and Azuaje
2003). This measures the average value of the similarity
between each cluster and its most similar cluster. A
lower DB index implies a better cluster configuration.
4. Calinski-Harabasz (CH) index (Dudoit and Fridlyand
2002). This measures between-cluster isolation and
within-cluster coherence. Its maximum value deter-
mines the optimal clustering configuration.
Another approach to validating the number of clusters
present in a dataset is to view clustering as a supervised
classification problem, in which we must also estimate the
‘‘true’’ class labels (Tibshirani and Walther 2005). Given the
output labels of a given clustering algorithm, we apply it to
train and build classification models (classifiers). Our goal is
to see how well the models can predict the labels using the
output of the clustering algorithms. The basic idea is that
‘true’ class labels will improve the prediction strength of the
classification models. Hence, the resulting ‘‘prediction
strength’’ measure assesses the quality of the clustering
results. We applied three different classification models
(support vector machines (SVM) (Burges 1998), neural
networks multilayer perceptron (MLP) (Jain et al. 1996), and
random forest (RF) (Breiman 2001)) to the clustering results.
An essential aspect of all cluster analysis is feature
selection/extraction. Using a large number of features (171 in
our case) increases the likelihood of feature redundancy. The
goal of feature selection is to remove irrelevant/redundant
features by finding the minimal feature subset necessary and
sufficient to support the target concept (Dash and Liu 1997).
The feature subset should improve and not degrade predic-
tion accuracy and be a fairly accurate representation of the
original feature distribution. To determine which facial
features were significant and discriminant among the 171
features, we applied three feature selection methods. parallel
scatter search algorithm (Garcı
´aLo
´pez et al. 2006), best first
search (Xu et al. 1988), and linear forward selection (Gutlein
et al. 2009). We validated the significance of the features by
reapplying the classification models. We expected that the
discriminant features would improve the prediction strength
of the models or at least not degrade performance of the
classifiers.
Statistical Comparisons
To determine significance of results obtained for the facially
defined clusters, we evaluated the statistical differences
between the clusters using the univariate one-way analysis of
variance (ANOVA) test along with the Student’s ttest for
continuous variables, v
2
and test for categorical variables.
The ANOVA test generalizes the Student’s ttest for between
comparisons for multiple groups. Hence, in addition, we
performed the student’s ttest for each distance measure for
comparisons between each pair of clusters to gain insight
into the significance of difference, where needed. The ttest
informs us on which pairs of clusters are actually statistically
different since the ANOVA’s pvalue only indicates that at
least one cluster is statistically different from another.
Results
Choosing Optimal Set of Clusters
The EM algorithm (unlike K-means, SOM and PAM) can
decide how many clusters to create by cross validation
based on resampling (Fraley and Raftery 2000). Thus we
ran the EM algorithm initially in this manner and it
Table 1 Evaluation of clustering algorithms using internal criteria
cluster validation measures
Clustering algorithm
(no. of clusters)
a
Cluster validation measures (index
scores)
Silhouette Davies-
Bouldin
Calinski-
Harabasz
Dunn
K-means (3) 0.12 1.65 12.42 0.84
K-means (4) 0.13 1.80 11.26 0.85
Expectation
maximization (3)
0.12 1.91 12.21 0.81
Self-organizing
feature map (4)
0.11 1.71 10.24 0.87
Self-organizing
feature map (3)
0.13 1.88 12.41 0.85
Partitioning around
medoids (3)
0.10 1.73 11.41 0.77
Best method according to each index is highlighted in bold
For all validation measures except Davies-Bouldin (DB), a higher
score indicates better cluster configuration. For the DB index, a lower
score implies better cluster configuration
a
Number of clusters in algorithm output result. For example,
K-Means (3) =3 cluster K-means result
J Autism Dev Disord
123
estimated 3 clusters within the dataset. We also reran the
EM algorithm with different values of k (number of clus-
ters) from 2 to 7. (We did not go beyond 7 due to the
limited size of the ASD data.) The best EM result (as
determined by cluster validity indices) was for k =3.
Based on this, for the remaining three cluster algorithms
we varied k from 3 to 7. The best EM result was compared
to the 21 outputs from the other three algorithms (K-means,
SOM, PAM). In Table 1, we compare the top six best
results. Based on the internal criteria cluster validation
indices, we selected the K-means output with k =3 as the
optimal cluster configuration. Those 3 clusters identified
within the 62 subjects ASD dataset were designated Cluster
1 (29 %, 18 boys), Cluster 2 (23 %, 14 boys), and Cluster 3
(48 %, 30 boys).
Machine learning techniques and evaluation metrics
were employed to verify the distinctness of the clusters by
training and testing three different classification models
(SVM, MLP and RF). Models were trained to classify
using the entire set of 171 facial geodesic distance mea-
sures. Given the limited size of the dataset, a threefold
cross-validation approach, that splits the dataset into 3
groups, was used. Thus, we train on two-thirds of the data
and test on the reminder third. Results are average of the
three separate runs (folds). Evaluation metrics used were
Classification Accuracy, Precision (Positive Predictive
Value), and Recall (Sensitivity). Classification accuracy is
defined as the percentage of test set samples that are cor-
rectly classified by the model. Precision (exactness) mea-
sures the proportion of actual positives that are correctly
identified by the model. Recall (completeness) is the ratio
of correctly classified samples to total number of samples
for a given class. We report both metrics to present a
complete depiction of the overall performance of the
models in terms of how precisely and completely it cor-
rectly identified each cluster on the average.
A minimal set of 31 features was derived by taking the
mathematical union of output results from three feature
selection algorithms. Using this minimal set of 31 facial
distance measures resulted in improved performance for
two models (SVM, RF), and equal performance for MLP
(Table 2). This validates that 31 features provide a robust
and discriminant representation of the entire 171 facial
distance measures.
To obtain a visual description of cluster separation, we
performed a Principal Component Analysis (PCA) on the
31 significant features. The distribution of the clusters
using the first two principal component axes is shown in
Fig. 2a. An illustration of the data using a dendrogram
based on mean linkage is shown in Fig. 2b. Note that the
dendrogram is only for visualization not interpretation of
data, as hierarchical clustering methods were not applied to
the data. Figure 3shows the distribution of the clusters plus
the control group. The control group overlaps strongly with
the cluster 3, partially with cluster 1and not at all with
cluster 2.
Facial Features Selection
The discriminant set of 31 facial geodesic features is
illustrated in Fig. 4. Each feature is denoted as the distance
from one anthropometric facial landmark to another. For
instance, BS feature indicates the distance measure from
the nasion (i.e. the midpoint of the forehead) to the
gnathion (i.e. the chin point).
All 31 distance measures were also verified to be sig-
nificant (pvalue less than 0.05) by doing a between com-
parison among all the clusters using the ANOVA test. In
addition, we performed the student’s ttest for each distance
measure for comparison between each pair of clusters. We
were interested in identifying which means were statisti-
cally different among all possible pairs of the three clusters
(clusters 1:2; clusters 2:3, clusters 3:1). 12 facial distance
measures were ascertained as statistically significant
among all three clusters, based on which features had a p
value of less than 0.05 for the pair-wise student’s ttests.
We describe these 12 facial distances, which were infor-
mative for all three clusters, in detail by mean and standard
Table 2 Evaluation of cluster separateness using all 171 facial distance features versus minimal 31 feature set in three classification models
Classification models Classification
accuracy (%)
Overall precision/
recall
Precision/recall per cluster
All 31 All 31 Cluster 1 Cluster 2 Cluster 3
All 31 All 31 All 31
Support vector
machine (SVM)
91.94 95.16 0.92/0.92 0.96/0.95 0.90/0.93 0.91/1.0 0.92/0.86 1.0/1.0 0.94/0.94 1.0/0.83
Neural networks multilayer
perceptron (MLP)
93.55 93.55 0.94/0.94 0.94/0.94 0.91/0.97 0.91/0.97 1.0/0.86 1.0/1.0 0.94/0.94 0.93/0.83
Random forest 88.71 91.94 0.91/0.89 0.92/0.92 0.81/1.0 0.88/0.97 1.0/0.79 1.0/0.86 1.0/0.78 0.94/0.89
Overall performance of the classfication models improved when trained and tested using minimal set of 31 features rather than all 171 features,
except in the case of cluster 3 for SVM and MLP
J Autism Dev Disord
123
deviation values (Table 3). Clustering results presented in
Fig. 3along with Table 3validates cluster 2 group as a
very compact and separate group among the ASD study
population and the typically developing boys (the control
group) using facial geodesic distance measures.
For each cluster, we identified which set of features were
discriminant and useful in describing each cluster facially
(Fig. 5). Cluster 1 is described by overall decreased surface
facial heights(BS, ES, GS, HS, LS, MS, QS), combined with a
broader maxillary midface from the temporal landmark to the
lower nose landmarks (JK, JL). However, these individuals,
demonstrate some overlap with the typically developing
controls (Fig. 3). Interestingly, cluster 1 has the lowest stan-
dard deviation values for 7 of the 12 discriminant facial dis-
tances measurements (Table 3). This further verifies its
compactness as the relatively low standard deviation values
imply facial distances measured do not vary widely among the
group (Fig. 2a). Cluster 2 subjects are facially defined by
overall increased surface facial heights (BS, ES, GS, HS, LS,
MS, QS), a decreased mid-face height (HS), and longest
Fig. 2 Visualization of ASD clusters using aprincipal coordinates
analysis plot of eigenscores for the first two principal axes. Axis 1
accounts for 34.37 % of the variance within the entire sample, and
axis 2 accounts for 24.78 % of the variance. bDendrogram based on
mean linkage. Note: this is only for visualization not interpretation of
data, as hierarchical clustering methods were not applied to the data
Fig. 3 Visualization of ASD
clusters overlapped with the
control group of 36 boys using
principal coordinates analysis
plot of eigenscores for the first
two principal axes. Axis 1
accounts for 30.88 % of the
variance within the entire
sample, and axis 2 accounts for
24.58 % of the variance
J Autism Dev Disord
123
mouth widths (NO and NQ). They also show no overlap with
the control boys (Fig. 3). Cluster 2 is characterized by the
most exaggerated facial features (Table 3; Fig. 5) among the
ASD study population. For 11 of the 12 facial distance mea-
sures in Table 3, cluster 2 subjects either have the maximum
or the minimum distance among the three clusters. Cluster 3
appears to be in between clusters 1 and 2, based on facial
morphological features. They also have the smallest NO and
NQ surface distances. Similar to cluster 1, cluster 3 individ-
uals demonstrate considerable overlap with along with the
typically developing boys. Thus, we observe that majority of
the boys with ASD cluster with the typically developing
controls, as also demonstrated by Aldridge et al. (2011).
Clinical Results
The goal of this study section is to determine whether the
ASD subgroups defined by the cluster analysis are clini-
cally distinctive. The clinical phenotype associated with
each cluster is described based on five clinical areas. ASD
diagnostic measures (ASD subsets, ADI-R and ADOS
scores), outcome indicators (IQ, Adaptive Behavior, lan-
guage), neurologic indicators (head size, seizures, electro-
encephalogram and brain Magnetic Resonance Imaging
(MRI) results), family history (alcoholism, ASD symp-
toms), and clinical course (regression).
ASD Core Symptoms
ADI-R and ADOS scores, which indicate greater impair-
ment with higher scores, were above ASD diagnostic cut-
offs for subjects in each cluster affirming their autism
diagnoses (Table 4). Social dysfunction, measured by the
ADI-R, was most impaired in cluster 2, and significantly so
compared with cluster 3. Cluster 2 also contained the
highest percentage of nonverbal subjects; verbal subjects
were more impaired in clusters 1 and 2 than in cluster 3.
Repetitive behaviors were highest in clusters 1 and 2, with
cluster 1 having a statistical significance over cluster 3.
Consistent with the ADI-R, ADOS calculated severity
scores were higher for clusters 1 and 2. Overall, individuals
in cluster 2 were most impaired, though often not signifi-
cantly from cluster 1. Cluster 3 was less symptomatic
generally and with a wider range of scores suggesting a
more heterogeneous subset of individuals.
Intelligence and Adaptive Behavior Scores
Though long-term functional outcomes are difficult to
predict in ASD, IQ scores, language development and
adaptive functioning provide some direction (Table 5). All
intelligence scores (NVIQ, VIQ and FSIQ) indicate that
boys in cluster 2 have significantly lower intelligence than
those in either cluster 1 or 3. Cluster 1 presented the
highest scores throughout though differences were not
significantly different from those in cluster 3. Wide ranges
and high standard deviations indicate significant hetero-
geneity in IQ and adaptive functioning in the 3 clusters.
The Vineland II adaptive scores did not discriminate
between the three clusters to the same degree as IQ
(Table 6). Vineland Adaptive Scores were similar for the
three groups with the exception of lower communication
scores for cluster 2.
Clinical Course
A history of language regression at the onset of ASD
symptoms in the first 3 years occurred in cluster 2 subjects
more than twice as often as in clusters 1 or 3 (57.1 vs. 16.7
and 20 %). (Table 7). There was no significant difference
in language regression between clusters 1 and 3. When
regression history was compared with IQ there was a sig-
nificant inverse association for all intelligence scores such
that individuals whose ASD presented with regression had
the lowest IQ scores (Table 8).
ASD Behavioral Subtype Diagnoses
Though Autism behavioral subtype diagnoses are no longer
considered valid diagnostic indicators (Lord et al. 2012),
these data are available and do convey some information
about what the diagnosing clinicians thought about the
subjects. Cluster 2 boys consisted of 79 % Autistic Dis-
order, 14 % as PDD-NOS, and 7 % as Asperger Syndrome.
A key finding is that individuals in cluster 2 were
Fig. 4 Illustration of the 31 discriminant facial distance features for
ASD clusters. 31 minimal discriminant features set from feature
selection phase is illustrated on the face (BS, EJ, GJ, LS, CH, EN, GO,
MN, CK, EP, GS, MS, CL, EQ, HS, NO, CO, ES, JK, NQ, DH, FO, JL,
QR, DI, FP, JN, QS, DJ, FQ, and JQ). Though these distances are
described using straight lines, they are not straight but rather the
shortest lines along the surface from one landmark point to the other
J Autism Dev Disord
123
diagnosed primarily with Autistic Disorder (78.6 %)
whereas cluster 1 (50 % Autistic Disorder, 44 % Asperger
Syndrome, 6 % PDD-NOS) and cluster 3 (47 % Autistic
Disorder, 33 % Asperger Syndrome, 20 % PDD-NOS)
consist of a distribution of subtypes reflective of the total
study population (55 % Autistic Disorder, 31 % Asperger
Syndrome, 15 % PDD-NOS). Also, the Asperger diagnosis
is closely correlated with IQ measurements, especially
verbal IQ and verbal functioning. Separation of patients
proposed in this paper provides subsets of patients based on
a physical biomarker—facial morphology. Facially defined
clusters reflect separation between more severely autistic
children (previously grouped under Autistic Disorder) and
less severe (previously grouped in Asperger Syndrome and
PDD).
Neurologic Indicators
Complete data on neurologic indicators were available for
seizures and head circumference. Seizures were more
common in cluster 2 (28.6 %) than in clusters 1 (22.2 %)
or 3 (10.0 %) though differences were not statistically
different. This may reflect the small number and young age
of the subjects. Head size measured by orbital occipital
circumference and converted to Z scores for analysis,
revealed no significant differences. Cluster 1 had the
highest mean Z score (1.21) which was not statistically
different from clusters 2 (0.87) and 3 (0.70). This indicated
that the facial phenotypes were not driven by differences in
head size. Head size groups’ results for the facial distance
defined clusters also showed that clusters are not related to
macrocephaly, as the percentage of macrocephalic subjects
in each cluster were similar (cluster 1–28 %, cluster
2–29 %, and cluster 3–20 %, all–24 %) and not statisti-
cally significant.
Genetic Indicators
Genetic indicators are those data that may provide insight
into the genetic basis of ASD. These may include gender
Table 3 Statistically significant facial distance measurements across clusters
Landmark Indicates Facial description Cluster 1 Cluster 2 Cluster 3
Mean SD Mean SD Mean SD
BS
Nasion–gnathion
Facial height Mid nasal bridge to chin point ;2.05 0.05 :2.25 0.08 2.14 0.07
ES
Palpebrale inferius–gnathion
Facial height Rt Mid eye to chin point ;1.81 0.06 :2.00 0.08 1.92 0.08
GS
Endocanthion–gnathion
Facial height Lf Inner canthus to chin point ;1.87 0.05 :2.05 0.08 1.98 0.07
HS
Palpebrale inferius–gnathion
Facial height Lf Mid eye to chin point ;1.83 0.06 :2.01 0.09 1.93 0.07
FP
Endocanthion–labiale superius
Mid Face height Rt Inner canthus to mid upper lip 1.15 0.04 ;1.13 0.04 :1.21 0.03
JK
Frontotemporale–alare
Mid Face breadth Lf Lateral eye brow to Rt nasal edge :1.89 0.05 1.79 0.05 ;1.83 0.05
JL
Frontotemporale–pronasale
Mid Face breadth Lf lateral eye brow to nose septum :1.31 0.05 ;1.21 0.04 1.27 0.05
LS
Pronasale–gnathion
Lower Face height Nose septum to chin point ;1.30 0.06 :1.50 0.07 1.38 0.08
MS
Alare–gnathion
Lower Face height Lf lateral nose to chin point ;1.19 0.07 :1.42 0.09 1.26 0.08
QS
Crista philtri–gnathion
Lower face height Lf Cupids bow to chin point ;0.90 0.07 :1.15 0.09 0.96 0.08
NO
Cheilion–crista philtri
Mouth width Rt lateral mouth to Rt cupids bow 0.47 0.05 :0.53 0.06 ;0.43 0.04
NQ
Cheilion–crista philtri
Mouth width Rt lateral mouth to Lf cupids bow 0.65 0.05 :0.72 0.10 ;0.61 0.05
Significance of means of facial distances determined by univariate ANOVA test between the three clusters along with pairwise student ttest
(p\0.001)
SD standard deviation
J Autism Dev Disord
123
and family history of autism and related neuropsychiatric
disorders. Social Responsiveness Scale (SRS) and Broad
Autism Phenotype Questionnaire (BAPQ) scores which are
designed to assess the number of autism symptoms in the
parents of individuals with ASD were analyzed for the 42
SSC project boys. Though no significant differences were
found between the clusters (Table 9), it is noted that in
each of the three measures (SRS, BAPQ-Autism, BAPQ-
Traits), mothers of boys in cluster 2 had somewhat higher
scores, indicating possible genetic or epigenetic predispo-
sition to develop an ASD. The portion of parents with
alcoholism, which is known to be significantly higher than
in families identified through an ASD (Miles et al. 2003),
did not assort by cluster. Consistent with previously pub-
lished data (Constantino and Gruber 2005), (Sasson et al.
2013), the paternal BAPQ scores for the ASD study pop-
ulation were significantly higher compared to the maternal
scores (Table 9). The relationship to gender and other
neuropsychiatric disorders could not be measured since all
subjects were male and the SSC were precluded families
with significant histories of ASD or major neuropsychiatric
diagnoses. None of the subjects had a history of chromo-
somal or other autism related disorders.
Discussion
Children with ASD diagnoses comprise a heterogeneous
population with a wide range in type, number and severity
Table 4 ASD core symptoms distribution by cluster
Diagnostic measures Cluster
1 (18)
Cluster
2 (14)
Cluster
3 (30)
Social (ADI-R A) (cutoff =10)
Mean (SD) 23.27
(5.57)
25.58
(4.14)
19.80
(7.05)
Range 9–30 18–30 5–30
Three cluster comparison (pvalue) 0.02
Pairwise comparisons (Clusters
1:2, 2:3, 3:1)
(pvalue)
0.23 <0.01 0.09
Verbal scores (ADI-R B) (cutoff =8)
Mean (SD) 18.80
(3.19)
18.70
(2.50)
15.91
(5.06)
Range 14–24 16–23 7–23
Three cluster comparison (pvalue) 0.07
Nonverbal scores (ADI-R B) (cutoff =7)
Percent of group measured by
Nonverbal criteria
0.00 %
(0)
14.29 %
(2)
6.67 %
(2)
Repetitive behavior (ADI-R C) (cutoff =3)
Mean (SD) 8.87
(2.47)
7.75
(1.71)
6.80
(2.47)
Range 4–12 5–10 2–12
Three cluster comparison (p-value) 0.03
Pairwise comparisons (Clusters
1:2, 2:3, 3:1)
(pvalue)
0.18 0.18 0.02
ADOS calculated severity scores
Mean (SD) 7.47
(1.88)
7.55
(1.51)
6.52
(1.60)
Range 5–10 6–10 4–9
Three cluster comparison (pvalue) 0.15
Statistically significant p-values are highlighted in bold
ADI-R data was available for 83 % of both clusters 1 and 3 and 86 %
of cluster 2 while ADOS data was available for 83, 79 and 70 % of
data for clusters 1- 3 respectively
Significance figure derived using univariate ANOVA test between the
three clusters
Fig. 5 Illustration of statistically significant facial distance measure-
ments per cluster. aCluster 1: 2D representation. bCluster 1: 3D
facial surface distance description. cCluster 2: 2D representation.
dCluster 2: 3D facial surface distance description. eCluster 3: 2D
representation. fCluster 3: 3D facial surface distance description.
Note: Facial surface distance features are compared among the 3
clusters. Red lines indicate maximum, orange are minimum distances
while blue imply distance is neither maximum nor minimum among
the 3 clusters (Color figure online)
J Autism Dev Disord
123
of social deficits, behavior, communication, and cognitive
difficulties which undoubtedly reflect multiple etiologic
origins (Eaves et al. 1994). An initial step in search for
etiologically discrete autism subgroups is discovery of
phenotypic features that are present in some but not all
ASD subjects, relatively discrete, quantifiable and patho-
physiologically relevant (Miles 2011). We proposed that
facial morphology, assessed by Euclidean and Geodesic
distances between anatomical landmarks, could be used to
reveal biologic homogeneity within ASD. Aldridge et al.
(2011) showed that young boys diagnosed with ASD
project a distinctive facial phenotype compared to typical
controls. The ASD face was characterized by increased
breadth of the upper face, orbits and mouth, a flattener
nasal bridge and reduced height of the philtrum and max-
illary region. Moreover, their data suggested biologic
subsets that correlated with ASD severity.
Table 5 Intelligence scores by cluster
Outcome indicators Cluster 1
(18)
Cluster 2
(14)
Cluster 3
(30)
Full Scale IQ
a
Mean (SD) 95.1 (18.60) 69.8 (25.98) 86.5 (21.58)
Range 68–127 31–112 38–130
FSIQ \70 5.6 % (1) 42.9 % (6) 16.7 % (5)
FSIQ C70 72.2 % (13) 50.0 % (7) 63.3 % (19)
Three cluster
comparison (pvalue)
0.02
Pairwise comparisons
(Clusters 1:2, 2:3, 3:1)
(pvalue)
0.01 0.06 0.21
Verbal IQ
b
Mean (SD) 93.9 (20.68) 66.0 (29.66) 84.4 (26.50)
Range 65–121 13–112 23–126
VIQ \70 11.1 % (2) 42.9 % (6) 20.0 % (6)
VIQ C70 66.7 % (12) 50.0 % (7) 56.7 % (17)
Three cluster
comparison (pvalue)
0.02
Pairwise comparisons
(Clusters 1:2, 2:3, 3:1)
(pvalue)
0.01 0.08 0.23
Non verbal IQ
c
Mean (SD) 94.8 (16.13) 73.7 (26.66) 92.3 (18.66)
Range 70–129 33–119 53–129
NVIQ \70 0.0 % (0) 42.9 % (6) 10.0 % (3)
NVIQ C70 94.4 % (17) 50.0 % (7) 73.3 % (22)
Three cluster
comparison (pvalue)
0.01
Pairwise comparisons
(Clusters 1:2, 2:3, 3:1)
(pvalue)
0.02 0.04 0.64
Statistically significant p-values are highlighted in bold
Significance figure derived using univariate ANOVA test between the
three clusters
a
FSIQ scores were available for 78, 93, and 80 % of clusters 1–3
respectively
b
VIQ scores were available for 78, 93, and 77 % of clusters 1–3
respectively
c
NVIQ scores were available for 94, 93, and 83 % of clusters 1–3
respectively
Table 6 Vineland adaptive scores by cluster
Vineland II Scores Cluster 1
(18)
Cluster 2
(14)
Cluster 3
(30)
Vineland Composite Score
Mean (SD) 73.8 (11.6) 71.0 (8.10) 77.2 (9.98)
Range 57–95 56–84 56 -100
Three cluster
comparison (pvalue)
0.31
Communication
Mean (SD) 77.9 (10.79) 70.2 (8.83) 80.3 (10.95)
Range 57–98 54–81 57–103
Three cluster
comparison (pvalue)
0.04
Pairwise comparisons
(Clusters 1:2, 2:3, 3:1)
(pvalue)
0.07 0.01 0.55
Daily living skills
Mean (SD) 77.8 (14.29) 78.5 (13.90) 81.4 (13.66)
Range 59–101 62–109 58–117
Three cluster
comparison (pvalue)
0.73
Socialization
Mean (SD) 69.7 (12.83) 69.3 (8.27) 73.9 (10.19)
Range 50–91 48–80 54–96
Three cluster
comparison (pvalue)
0.39
Statistically significant p-values are highlighted in bold
Significance figure derived using univariate ANOVA test between the
three clusters
Vineland II scores were available for 67, 79, and 73 % of clusters 1–3
respectively. Vineland Composite scores were available for 67, 79,
and 67 % of clusters 1–3 respectively
Table 7 Language regression by cluster
Language regression Cluster
1 (18)
Cluster
2 (14)
Cluster
3 (30)
Total
(62)
Language regression %
(#)
16.7 %
(3)
57.1 %
(8)
20.0 %
(6)
27.4 %
(17)
Three cluster comparison
(pvalue)
0.02
Pairwise comparisons
(Clusters 1:2, 2:3, 3:1)
(pvalue)
0.02 0.02 0.24
Statistically significant p-values are highlighted in bold
Regression data was available for all subjects
Significance figure derived using v
2
test between the three clusters
J Autism Dev Disord
123
Our goals were to validate the Aldridge results and
identify mathematically stronger clusters using additional
statistical approaches. Identification of biologically valid,
clinically distinctive subgroups is expected to expedite the
search for autism genes and treatments. To minimize
ASD’s inherent heterogeneity, subjects were limited to
Caucasian prepubertal boys, aged 8 to 12 with no signifi-
cant dysmorphology or microcephaly. Facial distances
were measured and mapped from three-dimensional stereo-
photogrammetric images of these boys. Each of the
subjects was comprehensively evaluated for autism related
symptoms, neurologic, cognitive, familial and phenotypic
variants.
Three ASD subgroups were identified by cluster analysis
based on geodesic distances between facial landmarks
(Farkas 1994). Geodesic distance, defined as the shortest
surface distance between anatomical landmarks, has been
suggested as better suited to capture geometric structure of
3D models than Euclidean distance (Hamza and Krim
2006, Gilani et al. 2013). Our interpretation of the strength
of the cluster analysis was based on four well-known
internal criteria cluster validation indices (Silhouette,
Dunn, Davies-Bouldin, Calinski-Harabasz) (Table 1).
Cluster compactness is reflected by standard deviations
(Table 3; Fig. 2), and separation of the clusters from each
other is measured by prediction strength (as reflected by
classification accuracy, sensitivity, and positive predictive
value) of three classification models (Support Vector
Machine, Neural Networks Multilayer Perceptron, and
Random Forest) (Table 2; Fig. 3). Feature selection was
also performed using established techniques (parallel
scatter search algorithm, best first search, and linear for-
ward selection) to select a subset of 31 geodesic distances
that result in better classification and clustering of the data.
The three ASD subgroups, delineated by clusters 1, 2 and
3, have distinctive, though subtle, facial measurements.
Cluster 1 is described by a reduction in facial height mea-
sures, combined with broader maxillary midface defined by
temporal to lower nose landmarks conveys a shorter broader
face. Cluster 1 faces are well separated from clusters 2 and 3
(as illustrated by the Principal Component Analysis—
Fig. 2a); however, there is considerable overlap with typi-
cally developing subjects (Fig. 3). Features that describe
cluster 3 faces include a shorter mid-face breadth, quantified
by left lateral eye brow to right nasal edge, smaller mouth
width and a decreased distance from the temporal area on the
left to the outer edge of the right nasal alae, all of which
portray a narrow face. Cluster 3 also has some overlap with
the typical developing subjects.
Cluster 2 is mathematically the most distinctive and well-
defined cluster (Tables 2,3; Fig. 3). The faces are best
described by an increased facial height measurements along
the surface, with the exception of a shorter midface. Mouth
widths are also wider. (Tables 2,3; Fig. 3). Three supervised
learning models (Support Vector Machine, Neural Networks
Multilayer Perceptron, Random Forest) were used to verify
the classification accuracy of the three clusters (Table 2).
Using these models, we were able to almost perfectly train
Support Vector Machine Classifier and Multilayer Neural
Network Perceptron to identify cluster 2 correctly from the
minimal set of 31 facial measurements. An F-measure of 1.0
indicates perfect classification. Cluster 2 also does not
overlap with the control boys (Fig. 3). Thus, cluster 2
Table 8 Correlation between language regression and IQ
VIQ NVIQ FSIQ
Language regression (27.4 %, 17) 47.0 67.5 56.0
No language regression (72.6 %, 45) 90.0 93.9 91.6
pvalue* \0.01 \0.01 \0.001
* p value reported in each column is based on using v
2
test to com-
pare mean IQ scores of subjects with no language regression to those
that have
Table 9 Parental history of ASD symptoms and alcohol abuse by
cluster
Cluster1
(18)
Cluster2
(14)
Cluster3
(30)
pvalue*
Social Responsiveness Scale (SRS)
Mother [mean
(SD)]
27.7 (18.5) 38.6 (14.0) 30.1 (19.3) 0.31
Father [mean
(SD)]
34.4 (31.0) 29.1 (20.0) 27.0 (20.3) 0.70
Broad autism phenotype (BAPQ)—autism
Mother [mean
(SD)]
2.3 (0.9) 2.6 (0.9) 2.2 (0.9) 0.62
Father [mean
(SD)]
2.7 (0.7) 2.7 (0.9) 2.6 (1.1) 0.94
Broad autism phenotype (BAPQ)—traits
Mother [mean
(SD)]
81.3 (36.6) 88.5 (38.1) 81.0 (32.3) 0.84
Father [mean
(SD)]
95.8 (23.4) 98.8 (28.9) 95.2 (34.9) 0.95
Alcoholism
Mother
alcoholic
44.4 % (8) 42.9 % (6) 40.0 % (12)
Father alcoholic 50.0 % (9) 50.0 % (7) 50.0 % (15)
68 % of the subjects were enrolled in the Simons Simplex Collection,
thus they had no history of autism among 1st or 2nd degree relatives
and no close relatives with major neuropsychiatric disorders
Raw SRS and BAP scores were available for 67, 79, and 73 % of
Clusters 1–3 respectively
Alcoholism traits (parental history of alcohol abuse) data was avail-
able for all except 1 boy in cluster 3
* Significance figure derived using univariate ANOVA test between
the three clusters
J Autism Dev Disord
123
subjects not only show substantial cluster strength based on
compactness and separateness criteria within the ASD pop-
ulation but also is distinct from the typically developing
matched control group.
To determine whether these facial morphology based
clusters would identify analogous clinical or behavioral
subsets within the ASD diagnosis population, individuals
in each cluster were assessed, using standard measures for
ASD core symptoms, cognitive, adaptive, and language
skills, ASD subtype diagnoses, type of ASD onset and
parental autism broad phenotype indicators. Cluster 2
subjects demonstrate the most coherent clinical phenotype
with 79 % (11/14) described as Autistic Disorder, 14 % (2/
14) as PDD-NOS, and 7 % (1/14) as Asperger Syndrome.
They are clinically defined by significantly higher ADI-R
A (Social) scores, (which implies a severe social diagno-
sis), severe verbal scores and an overall highest ADI CSS
score. They also have the highest occurrence of non-verbal
patients (14 %), the lowest IQ and Vineland II adaptive
scores (except for daily living skills) in all categories. In
addition to greater severity on autism measures, cluster 2
boys had more seizures (28 %) than boys in clusters 1
(22 %) or 3 (10 %). Moreover, this subgroup reported a
likelihood of early language regression of 57 %, which is
more than twice the frequency reported for clusters 1
(17 %) and 3 (20 %). The association of frequent language
regression with cluster 2 and overall severity of Autistic
Disorder diagnosis of these subjects provide additional
evidence in line with Stefanotos’ prognosis (2008). Stef-
anotos’ findings suggest that the regressive subgroup of
children with ASD may differ from the congenital form of
the disorder in severity of behavioral symptoms and long-
term prognosis, although he argues that more evidence is
needed to justify them as a distinct subgroup with a dis-
tinguishable set of etiological considerations. Though sib-
ling data was not available, severity of the maternal SRS,
BAPQ—Autism, and BAPQ—Traits scores indicates an
underlying genetic etiology for the individuals in cluster 2.
Additional indicators of possible genetic differences
between the clusters, including gender, autism and other
psychiatric disorders in siblings and family members was
not available because 68 % of ASD study population was
from the Simons Simplex Collection (SSC). The SSC
project recruited ASD patients based on exclusion of
multiplex autism families and families with psychiatric
disorders in close family members.
Clinical phenotype of Cluster 1 subjects is described by
50 % (9/18) Autistic Disorder, 44 % (8/18) with Asperger
Syndrome, and 6 % (1/18) with PDD-NOS. They are
clinically defined by significantly higher (indicating greater
severity) ADI-R C (Repetitive Behavior) scores. They have
no occurrence of non-verbal patients along with the highest
IQ and are the least likely group to experience language
regression. Interestingly, this group also has the lowest
Vineland II adaptive daily living skills scores. Cluster 3
appears to represent the broad composition of children
diagnosed with ASD. This is the largest subgroup (48 %
(30/62)) with 47 % (14/30) of the boys described as
Autistic Disorder, 33 % (10/30) as Asperger Syndrome,
and 20 % as PDD-NOS. Clinically, they are defined by the
lowest ADI-R scores in all the categories, which implies
that this group has the least severe diagnosis socially,
verbally, and repetitive behavior wise. This group also has
the best Vineland II adaptive scores in all categories.
However, this group has lower IQ scores compared to
cluster 1 subjects, though much higher than cluster 2 sub-
jects. This may be due to the presence of 2 (6.7 %) non-
verbal boys in this group. There is a 20 % occurrence of
language regression in this group. It is important to
remember that both clusters 1 and 3 overlap with the
control boys. These two clusters are clinically distinct from
each other by their ADI-R scores with cluster 3 having
better scores than cluster 1. Table 10 provides a clinical
summary of each cluster in terms of the indicators/symp-
toms (autism core symptoms, cognitive, outcome, associ-
ated neurological symptoms and regression).
Our results are complimentary to previous study by
Aldridge et al. (2011) performed on a similar, overlapping
Table 10 Summary of clinical and behavioral severity levels for
each cluster
Clinical/behavioral
phenotypes
Cluster 1 Cluster 2 Cluster 3
Subjects
Social competency
(ADI-R)*
Severe Most
severe
Least
severe
Verbal/communication
(ADI-R, Vineland II)*
Severe Most
severe
Least
severe
Repetitive behavior
(ADI-R)*
Most
severe
Severe Least
severe
ASD severity (ADOS) Severe Most
severe
Least
severe
ASD diagnostic
subgroup (DSM-IV)
Asperger Autistic
disorder
PDD-NOS
Cognitive Level (VIQ,
NVIQ, FSIQ)*
Highest Lowest High
Language regression
(\year 3)*
Least
frequent
Most
frequent
Frequent
Parents
SRS—Mother Least
severe
Most
severe
Severe
BAPQ (autism)—Mother Severe Most
severe
Least
severe
BAPQ (traits)—Mother Severe Most
severe
Least
severe
* Comparison is significant, as determined by univariate ANOVA test
between the three clusters
J Autism Dev Disord
123
dataset (52 out of the 63 used previously in addition to 10
new boys) but with different research methodology. Key
methodology differences are geodesic rather than Euclid-
ean distance measurements, multiple clustering techniques
versus principal component analysis; and two additional
landmark points that further define measurements of facial
height. In this study, we base our cluster separation deci-
sion solely on the ASD group in contrast to Aldridge et al.,
which includes separation from the control group as part of
the cluster decision process. This report provides further
evidence that the cluster results are strong, with a high
degree of compactness and separateness not as easily
appreciated as in the initial study. It is gratifying that both
studies identified basically the same severe autism sub-
group (Cluster 2 or Subgroup 1); characterized by severe
ADI-R scores, low cognitive and functional IQ scores,
highest maternal SRS scores and significant language
regression. It is interesting to note that only 6 of the 12
boys identified by Aldridge et al. as belonging to the severe
autism group (Subgroup 1) were included in our current
study population. Based on our cluster analysis results, 5 of
these 6 boys were included in our severe autism group
(Cluster 2). Hence, both studies indicate that boys with
ASD have altered development of their facial structure. In
terms of Euclidean distance measurements, Aldridge et al.
describes the severe autism subgroup with a decreased
height of the facial midline and increased breadth of the
mouth as well as the length and height of the chin. It is
known that distance along the surface between two land-
mark points is not equivalent to the Euclidean distance
between these points. Our findings indicate that cluster 2,
our severe autism cluster, is characterized by an overall
increased facial surface height measurements (with the
exception of decreased mid-face height), and larger mouth
widths compared to the measurements in individuals in
clusters 1 and 3. This describes a longer face along the
surface. Both studies indicate that distance measurements
that describe decreased height of facial midline and long
mouth widths are key biological traits for the severe autism
group. This study demonstrates the generalization of facial
phenotypes as a viable biomarker for identifying ASD
subgroups, independent of measurement type (Euclidean
vs. Geodesic) or cluster technique.
Our findings also indicate a strong association between
language regression and cognitive performance, as indi-
cated by IQ scores of cluster 2 as well as the entire study
population. According to Table 7, 27 % of our study
population has experienced language regression, which is
consistent with the composition studied in literature about
language regression in ASD (Jones and Campbell 2010). A
pairwise comparison of the mean IQ scores in all three
categories (VIQ, NVIQ, and FSIQ) between the regressed
group and the non-regressed group (Table 8) reveals that
the regressed group has significantly much lower IQ scores.
Regression is a relatively common phenomenon in many
pediatric neurologic disorders and has been linked to
genetic diagnoses (Miles 2011). Though several reports
have suggested that the eventual outcome in children with
regression is that of a lower language level, lower IQ and
lower adaptive level compared with those who do not
regress, other studies have found no difference in outcome
(Baird et al. 2008). Baird et al. found children with broad
ASD diagnoses showed greater symptom severity in the
presence of some language regression versus no regression.
The outcomes from our study provide additional substan-
tiation in support of a statistical correlation between lan-
guage regression and cognitive performance.
Our findings also provide additional evidence that
macrocephaly is an independent autism specific feature of
autism. Head size results presented confirms that clusters
are not related to macrocephaly, which is a relatively non-
specific finding in autism (Miles et al. 2000). Importantly,
lack of association of head size with the clusters clearly
indicates that brain and head growth are not the cause of
the facial phenotypes.
The primary limitation of this study was the relatively
small size and lack of some clinical data. Primary strength
was the participation of mathematical and statistical sci-
entists who designed a statistical approach that confirmed
the validity of the cluster methodology.
Conclusion
Using comprehensive cluster analysis techniques, facial
surface measurements were investigated in a cohort of 62
eight to twelve year old boys with essential ASD. Our
results validate and extend the work of Aldridge et al.
(2011) which showed for the first time that facial mor-
phology differed significantly between groups of boys with
ASD and matched controls and that subsets with distinctive
facial morphology could be identified. Moreover, by using
similar but different clustering methods, we also identified
a comparable subset of boys with a classical autistic dis-
order phenotype characterized by lower IQs and Vineland
Adaptive behavior scores, severe autism symptoms mea-
sured by gold standard autism diagnostic measures (ADI-R
and ADOS), and more than twice likelihood of early lan-
guage regression.
Based on these two studies, we assert that facial struc-
ture, based on 31 geodesic facial distances, should be
considered a potentially useful biomarker to separate out a
biologically discrete and homogeneous ASD subset for
further study. This may help predict disorder severity and
regression and has translational relevance as this ASD
subset may represent genetically distinct individuals for
J Autism Dev Disord
123
whom specific treatment options may be tailored. Three
dimensional facial imaging, which can be acquired with
commercially available 3 dimensional systems already
located in many university based tertiary care hospitals,
should become a feasible autism biomarker with which to
delineate homogeneous populations.
References
Aldridge, K., George, I., Cole, K., Austin, J., Takahashi, T. N., Duan,
Y., et al. (2011). Facial phenotypes in subgroups of pre-pubertal
boys with autism spectrum disorders are correlated with clinical
phenotypes. Molecular Autism, 2(1), 15.
American Psychiatric Association. (2000). Diagnostic and statistical
manual of mental disorders (4th ed.). Washington, DC: APA.
American Psychiatric Association. (2013). Diagnostic and statistical
manual of mental disorders (5th ed.). Arlington, VA: American
Psychiatric Publishing.
Arbelaitz, O., Gurrutxaga, I., Muguerza, J., Pe
´rez, J. M., & Perona, I.
(2013). An extensive comparative study of cluster validity
indices. Pattern Recognition, 46(1), 243–256.
Baird, G., Charman, T., Pickles, A., Chandler, S., Loucas, T.,
Meldrum, D., et al. (2008). Regression, developmental trajectory
and associated problems in disorders in the autism spectrum: The
SNAP study. Journal of Autism Development Disorders, 38(10),
1827–1836.
Bolshakova, N., & Azuaje, F. (2003). Cluster validation techniques
for genome expression data. Signal Processing, 83(4), 825–833.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
Burges, C. J. (1998). A tutorial on support vector machines for pattern
recognition. Data Mining and Knowledge Discovery, 2, 121–167.
Constantino, J. N., & Gruber, C. P. (2005). Social Responsiveness
Scale. Los Angeles, CA: Western Psychological Services.
Dash, M., & Liu, H. (1997). Feature selection for classification.
Intelligent data analysis, 1(3), 131–156.
Dudoit, S., & Fridlyand, J. (2002). A prediction-based resampling
method for estimating the number of clusters in a dataset.
Genome Biology, 3(7), 1–21.
Eaves, L. C., Ho, H. H., & Eaves, D. M. (1994). Subtypes of autism
by cluster analysis. Journal of Autism and Developmental
Disorders, 24(1), 3–22.
Ewing, J. A. (1984). Detecting alcoholism: The CAGE questionnaire.
JAMA: Journal of the American Medical Association, 252,
1905–1907.
Farkas, L. G. (1994). Anthropometry of the head and face in clinical
practice. In L. G. Farkas (Ed.), Anthropometry of the head and
face (2nd ed., pp. 71–77). New York: Raven Press.
Farkas, L. G., & Posnick, J. C. (1992). Growth and development of
regional units in the head and face based on anthropometricmea-
surements. The Cleft Palate-Craniofacial Journal, 29(4), 301–302.
Fraley, C., & Raftery, A. E. (2000). Model-based clustering,
discriminant analysis, and density estimation. Journal of the
American Statistical Association, 97, 611–631.
Garcı
´aLo
´pez, F., Garcı
´a Torres, M., Melia
´Batista, B., Moreno Pe
´rez,
J. A., & Moreno-Vega, J. M. (2006). Solving feature subset
selection problem by a parallel scatter search. European Journal
of Operational Research, 169(2), 477–489.
Gilani, S. Z., Shafait, F., & Mian, A. (2013). Biologically significant
facial landmarks: How significant are they for gender classifica-
tion? In IEEE international conference on digital image comput-
ing: Techniques and Applications (DICTA) (pp. 1–8).
Gutlein, M., Frank, E., Hall, M., & Karwath, A. (2009). Large-scale
attribute selection using wrappers. IEEE symposium on compu-
tational intelligence and data mining CIDM’09,2009 (pp.
332–339).
Halkidi, M., Batistakis, Y., & Vazirgiannis, M. (2001). On clustering
validation techniques. Intelligent Information Systems Journal,
17(2–3), 107–145.
Hamza, A. B., & Krim, H. (2006, August). Geodesic matching of
triangulated surfaces. IEEE Transactions on Image Processing,
15(8), 2249–2258.
Han, J. C. (1990). Shortest paths on a polyhedron. Sixth annual
symposium on Computational geometry, SCG’90, (pp. 360–369).
Hartigan, J. A., & Wong, M. A. (1979). Algorithm AS 136: A
k-means clustering algorithm. Journal of the Royal Statistical
Society: Series C (Applied Statistics), 28(1), 100–108.
Jain, A. K. (2010). Data clustering: 50 Years beyond K-means.
Pattern Recognition Letters, 31(8), 651–666.
Jain, A. K., Jianchang, M., & Mohiuddin, K. M. (1996). Artificial
neural networks: a tutorial. Computer, 29(3), 31–44.
Jones, L. A., & Campbell, J. M. (2010). Clinical characteristics
associated with language regression for children with autism
spectrum disorders. Journal of Autism and Developmental
Disorders, 40(1), 54–62.
Kaufman, L., & Rousseeuw, P. J. (1990). Finding groups in data: An
introduction to cluster analysis. New York: Wiley.
Kohonen, T. (1998). The Self-organizing maps. Neurocomputing,
21(1), 1–6.
Kova
´cs, F., Lega
´ny, C., & Babos, A. (2005). Cluster validity
measurement techniques. 6th International symposium of hun-
garian researchers on computational intelligence.
Lord, C., Petkova, E., Hus, V., Gan, W., Lu, F., Martin, D. M., et al.
(2012). A multisite study of the clinical diagnosis of different
autism spectrum disorders. Archives of General Psychiatry,
69(3), 306–313.
Lord, C., Risi, S., Lambrecht, L., Cook, E. H., Leventhal, B. L.,
DiLavore, P. C, et al. (2000, June). The autism diagnostic
observation schedule-generic: A standard measure of social and
communication deficits associated with the spectrum of autism.
Journal of Autism and Developmental Disorders, 30(3),
205–223.
Lord, C., Rutter, M., Goode, S., Heemsbergen, J., Jordan, H.,
Mawhood, L., et al. (1989). Autism diagnostic observation
schedule: A standardized observation of communicative and
social behavior. Journal Autism Development Disorder, 19,
185–212.
Lord, C., Rutter, M., & Le Couteur, A. (1994). Autism Diagnostic
interview-revised: A revised version of a diagnostic interview for
caregivers of individuals with possible pervasive developmental
disorders. Journal of Autism Development Disorder, 24,
659–685.
Miles, J. H. (2011). Autism subgroups from a medical genetics
perspective. In Autism spectrum disorders (pp. 705–721).
Oxford: Oxford University Press.
Miles, J. H., Hadden, L. L., Takahashi, T. N., & Hillman, R. E.
(2000). Head circumference is an independent clinical finding
associated with autism. American Journal of Medical Genetics,
95(4), 339–350.
Miles, J. H., Takahashi, T. N., Haber, A., & Hadden, L. (2003).
Autism families with a high incidence of alcoholism. Journal of
Autism and Developmental Disorders, 33(4), 403–415.
Miles, J. H., Takahashi, T. N., Hong, J., Munden, N., Flournoy, N.,
Braddock, S. R., et al. (2008). Development and validation of a
measure of dysmorphology: Useful for autism subgroup classi-
fication. American Journal of Medical Genetics Part A, 146A,
1101–1116.
J Autism Dev Disord
123
Mitchell, J. S., Mount, D. M., & Papadimitriou, C. H. (1987). The
discrete geodesic problem. SIAM Journal Computing, 16(4),
647–668.
Quynh, D., He, Y., Xin, S.-Q., & Chen, Z. (2012). An intrinsic
algorithm to compute geodesic distance fields on triangle meshes
with holes, graphical models. Proceedings of Geometric
Modeling and Processing GMP’12, 74(4), 209–220.
Rousseeuw, P. (1987). Silhouettes: A Graphical Aid to the Interpre-
tation and Validation of Cluster Analysis. Computational and
Applied Mathematics, 20, 53–65.
Sasson, N. J., Lam, K. S., Parlier, M., Daniels, J. L., & Piven, J.
(2013). Autism and the broad autism phenotype: Familial
patterns and intergenerational transmission. Journal of Neuro-
developmental Disorders, 5(11).
Sparrow, S., Balla, D., & Cicchetti, D. (1984). Vineland Adaptive
Behavior Scales. Circle Pines, MN: American Guidance Service.
Stefanatos, G. A. (2008). Regression in autistic spectrum disorders.
Neuropsychology Review, 18(4), 305–319.
Tibshirani, R., & Walther, G. (2005). Cluster validation by prediction
strength. Journal of Computational and Graphical Statistics,
14(3), 511–528.
Wade, J. L., Cox, N. B., Reeve, R. E., & Hull, M. (2014). Brief report:
Impact of child problem behaviors and parental broad autism
phenotype traits on substance use among parents of children with
ASD. Journal of Autism and Developmental Disorders, 1–7.
Wang, K., Wang, B., & Peng, L. (2009). CVAP: Validation for cluster
analyses. Data Science Journal, 8, 88–93.
Witten, I. H., & Frank, E. (2005). Data mining: Practical machine
learning tools and techniques (2nd ed.). Los Altos, CA: Morgan
Kaufmann.
Xin, S.-Q., Quynh, D., Ying, X., & He, Y. (2012). A global algorithm
to compute defect-tolerant geodesic distance. In ACM SIG-
GRAPH ASIA 2012 Technical Briefs, pp. 1–23.
Xu, R., & Wunsch, D. (2005). Survey of clustering algorithms. IEEE
Transactions on Neural Networks, 16(3), 645–678.
Xu, L., Yan, P., & Chang, T. (1988). Best first strategy for feature
selection. In Proceedings of ninth international conference on
pattern recognition (pp. 706–708).
J Autism Dev Disord
123
