DATA Mining of Clinical Variables and Biological Endophenotypes in Autistic Patients Using Fourth Generation Artificial Neural Networks

Background:  Several studies have attempted to partition autistic individuals into subtypes ideally homogeneous in terms of clinical presentation and/or underlying pathogenesis. Clinical subtyping has been defined one of the major short-term challenges in child and adolescent psychiatry. This is especially true for autism research, since clinical heterogeneity represents one of the hallmarks of ASD. We have recently analyzed the autistic phenotype taking into account observable behaviors, patient- and family-history variables, and biological endophenotypes. Using principal component and cluster analysis on 245 patients, we previously described at least four principal components and four patient clusters (Sacco et al., Autism Res. 2010 and 2012).

Objectives:  To identify specific patterns linking biological endophenotypes, such as macrocephaly and elevated serotonin blood levels, to autism clinical profiles.

Methods:  Artificial Neural Network were applied to a complete data set of 110 ASD patients encompassing 25 variables spanning clinical features, family history, morphological and biochemical quantitative traits. We applied semantic connectivity maps (AutoCM), a fourth generation artificial neural network able to detect non-linear trends and associations among variables with significantly greater power as compared to the traditional parametric statistics employed in our previous study. The matrix of connections, visualized through the minimum spanning tree, maintains non-linear associations among variables and captures schemes among clusters of variables. The strength of association in semantic connectivity maps ranges from 0 to 1 (i.e., from no to full association).

Results:  [1] clinical variables tend to cluster around two configurations: (a) “lower functioning”, which has its central node in the presence of motor stereotypies, strongly connected with intellectual disability (0.99), verbal stereotypies (0.98), hyperactivity (0.98), reduced pain sensitivity (0.97), history of regression (0.94) and self-injurious behaviors (0.93); (b) “higher functioning”, which has its two central nodes in positive history of allergies or immune disease in the patient, or in first-degree relatives, tightly linked to each other (0.97) and with obstetric complications (0.97), delayed onset of social smile (0.97), presence of any infectious disease at autism onset (0.96), pre-term delivery (0.89), normal intellectual level (0.89), and a DSM-IV Asperger (0.88) or PDD-NOS (0.85) diagnosis. [2] Macrocephaly is associated with a positive history of allergy and immune disease in first-degree relatives (0.92) and to a lesser extent with muscle hypotonia (0.77). [3] Hyperserotoninemia may be connected with abnormal EEG pattern and/or history of seizures in males (0.80), whereas in females it appears linked to positive history of allergy/immune disease in first-degree relatives and muscle hypotonia, although sample size limitations for females do not yet allow reliable coefficient estimations.

Conclusions:  

AutoCM algorithms show several complex patterns which replicate and largely extend previous findings obtained with parametric approaches. New insights, such as those possibly linking hyperserotoninemia with abnormal EEG patterns, if replicated may allow novel hypothesis generation and experimental designs. These results will be replicated in an independent sample, so as to better define the relationship between biological endophenotypes, biomarkers and clinical features involved in autism.