Autism diagnosis is typically based on clinical observation and assessment, a process that can be complex and subjective. To refine this approach, researchers leveraged a large language model (LLM) to determine which behaviours and observations are most strongly associated with an autism diagnosis.
Their results show that repetitive behaviours, special interests, and perception-based behaviors are most strongly associated with an autism diagnosis. These findings have the potential to improve diagnostic guidelines by reducing the emphasis on social factors, which are the primary focus of the DSM-5 guidelines but were not classified by the model as the most relevant in diagnosing autism.
“Our goal was not to suggest that AI tools could replace clinicians for diagnosis,” says senior author Danilo Bzdok of the Mila Quebec Artificial Intelligence Institute and McGill University in Montreal. “Rather, we sought to quantitatively define exactly which aspects of observed behavior or patient history a clinician uses to reach a final diagnostic determination. In doing so, we hope to empower clinicians with diagnostic instruments that better align with their empirical realities.”
The researchers leveraged a transformer-based language model pre-trained on approximately 489 million unique sentences. They then fine-tuned the LLM to predict diagnostic outcomes using a collection of more than 4,000 reports written by clinicians assessing patients for autism.
These reports—often reviewed by multiple clinicians—contained accounts of observed behavior and relevant patient history but did not include a suggested diagnostic outcome. The team developed a bespoke LLM module that identified specific sentences within the reports that were most relevant to an accurate diagnosis prediction.
Next, they extracted numerical representations of these highly autism-relevant sentences and compared them directly with the established diagnostic criteria outlined in the DSM-5. The researchers were surprised by how clearly the LLM distinguished the most diagnostically relevant criteria. Their framework highlighted that repetitive behaviors, special interests, and perception-based behaviors were the most significant indicators of autism.
While these criteria are considered in clinical settings, current diagnostic guidelines place greater emphasis on deficits in social interaction and communication skills.
The team expects that their framework will be useful for researchers and medical professionals working with various psychiatric, mental health, and neurodevelopmental disorders, where clinical judgment plays a central role in diagnosis.
ANI
