Project description

The Autism Spectrum (AS) is a neurodevelopmental condition characterized by socio-communicative peculiarities and repetitive and stereotyped behaviors. The main tool used to diagnose AS is the DSM-5 classification manual. To date, the scientific community has widely accepted the existence of great heterogeneity in the AS phenotype, however we hypothesize that this heterogeneity could be at least partially caused by the way DSM-5 criteria are built and used. In support of this hypothesis, it has been shown that the presentation of “frank” autism is recognized intuitively in less than 10 minutes by experienced clinicians. Clinicians may recognize “frank” autism according to qualitative criteria, some of which are not included in the DSM-5 (atypical prosody for example), with an interjuge agreement greater than 80%. Beyond the use of DSM-5 criteria, diagnosticians therefore base their diagnostic decision-making on a complex implicit process involving subjective phenotypic markers of autism recognition, in addition to the explicit verification of the presence or lack of signs. These markers remain to this day still unknown. Our project aims to use different sources rich in clinical descriptions of autism (international databases, clinical records) to which we will apply machine learning strategies and language processing algorithms to reveal these subjective markers. Our results will allow us to objectively define the behavioral and phenotypic signatures of subjectively perceived autism, and thus to refine autism diagnostic criteria to improve detection of this condition.

For research purposes, this will make it possible to identify a “prototypical” subpopulation presenting almost all of these concretely defined signs. From this population, researchers will be able to search for markers with less noise than in the population as currently defined.