Leggi in italiano
The support needed by autistic people can vary a lot, particularly at an early age when interventions can be very effective due to brain plasticity. Credit: Maria Dubova/ iStock / Getty Images Plus. 
A group of researchers at the Italian Institute of Technology (IIT) has developed an algorithm1 that identifies two main subtypes of autistic children starting from the Vineland Adaptive Behavior Scales (VABS), a standard interview used to measure adaptive behaviors in children and adults with intellectual disabilities and neurodivergent conditions. VABS covers communication, daily living skills, socialization, and motor skills. The two subtypes, one with lower and the other with higher scores on average, are predictive of the adaptive skills children will have later in life.
“Our algorithm could be employed to facilitate future research evaluating the efficacy of interventions and treatments”, says Michael Lombardo, who led the study and heads the Laboratory for Autism and Neurodevelopmental Disorders at IIT. Lombardo thinks this could also help to identify biological traits linked to different clinical outcomes.
Researchers used the VABS scores made publicly available by the US National Institutes of Mental Health. They constructed two datasets, the first one comprising nearly 1,000 children below six years old and the second one made by nearly 2,000 individuals between 6 and 61 years old. They then looked for clusters in the data.
“We developed a clustering algorithm that is highly generalizable and replicable”, says Veronica Mandelli, a researcher at IIT and first author of the study. The algorithm, called reval, starts by splitting the dataset into a training set and a validation set, and then further divides the training set into two subsamples. It then tries several strategies to classify individuals into clusters, and chooses the one that maximizes the agreement between the clusters in the two samples. Then it does the same thing for the validation set. After several iterations, the algorithm selects the clustering that is most stable across all samples.
For children below six years old, scientists found three equally populated autism subtypes that they called ‘low’, ‘medium’ and ‘high’, since in each domain considered by the VABS the scores spread around three distinct and growing mean values. In the older group, there are still three subtypes but one of them comprises fewer than 3% of subjects with very low scores, while the other two are nearly of the same size and include individuals with relatively high and low scores.
The researchers find that subjects in the early ‘high’ or ‘low’ subtype will probably remain in that same ‘high’ or ‘low’ subtype at older ages. The early ‘medium’ subtype is much more ambiguous with respect to laters outcomes.
The group developed a free web-based application to allow other scientists to run the algorithm on their own databases. The authors hope that this will encourage retrospective analyses of clinical trials data.
“We see substantial inter-individual differences in how autistic children respond to early behavioral interventions”, says Liliana Ruta, child neuropsychiatrist at the Institute for biomedical research and innovation of the National Research Council, in Messina. ”This algorithm could help us tailor support to maximize outcome”, she adds. Ruta observes that this work considered people that received different support during their life. “It would be interesting to run it on more homogeneous cohorts with respect to this aspect”, she concludes.
