From Sheldon Cooper to Sherlock Holmes, autism-coded characters on screen have long been defined by their wacky social skills, constantly misunderstanding jokes and saying things that others find rude.
When it comes to diagnosing real people with autism, professionals often focus on social skills too – but this emphasis might be misplaced.
A recent study concluded that other autistic traits are a better predictor of whether someone will be diagnosed with the condition or not – traits such as repetitive behaviours, special interests and sensory differences.
“We think that our research has the potential for large impact,” study co-author Jack Stanley, PhD student in biochemistry and machine learning at McGill University, Montreal, told BBC Science Focus.
“From the perspective of the autism community, we think that this study may encourage the revision or re-weighting of longstanding clinical criteria for diagnosing autism.”
People with autism are generally diagnosed through a clinical observation, where a health professional assesses the person based on a list of potential traits, labelled as autistic – but this process relies largely on subjective judgement and intuition.
Researchers at McGill wanted to understand how clinicians decide that a person is autistic in a more quantitative way, so they used a large language model (LLM): a type of artificial intelligence that is designed to understand and analyse human text.
The scientists fed an LLM more than 4,000 reports, written by clinicians who were assessing patients for autism, and trained the machine to accurately predict whether each person would be diagnosed with the condition.
“But we went a step further,” said Stanley. “We did not simply want to construct an ‘autism detector,’ nor did we aim to replace clinicians with an LLM.
“Instead, we wanted to empower clinicians to better understand the most relevant factors to look for when diagnosing a patient with autism.”
So, the scientists trained the LLM to spot specific sentences in the reports that were most relevant to predicting a diagnosis, and rate them based on their importance.
When the scientists compared the LLM’s findings with what Stanley called the “trusted gold standard criteria for autism” – the American Diagnostic and Statistical Manual of Mental Disorders (DSM-5) – they found that social skills were much less important than repetitive behaviours and special interests.
“This is in contrast to some 40 years of research and the current clinical guidelines,” said Stanley.
This research did not distinguish between male and female patients, so these findings may be more or less relevant for individuals of different genders.
The scientists hope their findings will help medical professionals reevaluate what is most relevant when diagnosing someone with autism.
Read more:
- No, it’s not just a trend: The real reasons adult ADHD and autism diagnoses are on the rise
- Is autism found in any other animals?
- Autism could soon be detected in your poop. Here’s why
About our expert:
Jack Stanley is a PhD student at McGill University, Montreal and Mila AI Institute, where he is studying how machine learning can be applied to biological and medical problems. Prior to McGill, he completed an Honours BSc in statistics and biochemistry at the University of Toronto.
