Objectives: To identify clusters of ASD patients characterized by specific patterns of principal components and to estimate their frequency.
Methods: We performed a k-means clustering of the same 245 patients assessed in our principal component analysis, using regression-based factors, each representing one cumulative component score. The K-means method is an unsupervised learning algorithm that assumes k clusters fixed a priori and defines k centroids, one for each cluster.
Results: ASD patients could be categorized into four clusters: (a) 43 patients (17.6%) have intense immune-related symptoms, accompanied by circadian and sensory issues; (b) 44 patients (18.0%) display intense circadian and sensory symptoms, with some developmental delay and stereotypies, but little or no immune dysfunction; (c) verbal and/or motor stereotypic behaviors predominate over the other three components in 73 (31.0%) patients, and (d) 83 (33.9%) patients show a mixture of all four components, with developmental delay somewhat more pronounced, and the other components less pronounced, compared to the remaining three clusters. The “immune” component II exerts the highest discriminatory power (F=111.247, df 241, P=2.7 x 10-45), followed by “stereotypic” component IV (F=98.165, df 241, P=1.4 x 10-41).
Conclusions: Despite the long-recognized interindividual variability in clinical phenotype, it seems possible to begin dissecting clusters of autistic patients based on a set of clinical, patient and family history variables. We are in the process of replicating and extending these results using a neural network approach. If replicated, these clusters may be used to explore differences in genetic underpinnings, disease course and severity, as well as in response to therapies.