An analysis of digital medical records using large language models (LLMs) is challenging a deeply held belief about the clinical characteristics of autism.
A recent study conducted by scientists at The Neuro (Montreal Neurological Institute-Hospital) of McGill University and Mila (Quebec Artificial Intelligence Institute) has shown that factors related to social communication may not be as indicative of this disorder as previously believed.
This finding challenges the standard diagnostic method for autism, which involves an assessment based on reference manuals such as the
Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5). The DSM-5 lists two categories of diagnostic criteria for autism: one relating to behaviors, sensitivities and interests, and another focusing on differences in the sphere of communication and social interactions.
For the study, scientists customized an artificial intelligence (AI) model to analyze over 4,200 clinical reports of children in Quebec. The analysis results showed that criteria related to socialization, such as emotional reciprocity, nonverbal communication and relationship building, were not closely linked to an autism diagnosis.
In other words, these criteria were not significantly more present in individuals diagnosed with autism than in those where this diagnosis was ruled out. On the other hand, criteria related to repetitive motor movements, hyperfixation on certain interests, and unusual sensitivity to sensory stimuli were strongly associated with an autism diagnosis.
In light of these results, published in the journal
Cell, scientists suggest that the medical community may need to reconsider the importance it places on current criteria and focus more on repetitive behaviors and special interests.
The potential of AI for rapid and accurate diagnostics
Currently, establishing an autism diagnosis relies on clinical evaluation; there is no biological test to analyze genes, blood or brain images. This is a lengthy process that can delay access to essential support services. According to researchers, diagnosis could be accelerated and made more accurate by focusing on the most predictive characteristics of autism. They highlight AI's potential to refine this process.
"Large language model technology might one day lead us to rethink our definition of autism," notes Danilo Bzdok, lead author and scientist at The Neuro and Mila. "This data-driven re-examination of current neurological disorders complements work that has traditionally been done entirely by expert groups and human judgment."
This study was funded by Brain Canada Foundation, Health Canada, the National Institutes of Health, the Canadian Institutes of Health Research, the Healthy Brain for Healthy Life program, the Canada Excellence Research Chairs program, and the Canadian Institute for Advanced Research.
The study
The article "Language models deconstruct the clinical intuition behind diagnosing autism" by Jack Stanley, Emmett Rabot, Siva Reddy, Eugene Belilovsky, Laurent Mottron and Danilo Bzdok was published in the journal
Cell.