Artificial intelligence could help experts identify toddlers who may be autistic, researchers say. They have developed a screening system that they say has an accuracy of around 80% for children under the age of two.

The researchers say their approach, which relies on a type of AI called machine learning, could provide benefits.

Dr Kristiina Tammimies, co-author of the study from the Karolinska Institutet in Sweden, said: “By using the AI ​​model, it may be possible to use available information and identify individuals at increased risk of autism earlier, so they can receive diagnosis and help earlier.”

But, she added: “I want to emphasize that the algorithm cannot diagnose autism, as this should (still) be done with gold standard clinical methods.”

This is not the first time that researchers have used AI to detect autism. Scientists have previously used this technology in combination with retinal scans of children.

Tammimies and colleagues write in the journal Jama Network Open about how they collected data from a U.S. research initiative called the Spark Study. This data comes from 15,330 children with an autism diagnosis and 15,330 children without.

The team describes how they focused on 28 measures that could be easily taken before children were 24 months old, based on parent-provided information from medical and background questionnaires, such as age at first smile.

They then created machine learning models that looked for different patterns in combinations of these features in children with autism and children without autism.

After using the data to build, tune, and test four different models, the team chose the most promising model and tested it on a further dataset of 11,936 participants for whom data on the same characteristics was available. In total, 10,476 of these participants had an autism diagnosis.

Results show that the model correctly identified a total of 9,417 (78.9%) participants with or without autism spectrum disorder, with an accuracy of 78.5% for children aged 0 to 2 years, 84.2% for children aged 2 to 4 years, and 79.2% for children aged 4 to 10 years.

Further testing with another set of data from 2,854 individuals with autism found that the model correctly identified 68% of individuals with such a diagnosis.

Tammimies said: “This dataset was another research cohort with families with only one child with autism and some parameters were missing, which made the performance a little bit lower, which shows we still need some more development.”

According to the researchers, the factors that mattered most to the model’s predictions included difficulty eating food, the age at which children were first able to form longer sentences, the age at which they were toilet trained, and the age at which they first laughed.

The team added that an additional analysis comparing participants the model correctly identified as having autism with participants who were incorrectly identified as non-autistic suggested the model tended to identify autism in individuals with more severe symptoms and more general developmental problems.

However, some experts cautioned against this, noting that the model’s ability to correctly identify people without autism was only 80%, meaning that 20% would have been incorrectly flagged as at risk for autism. They also noted that pushing for early diagnosis could be problematic.

According to Professor Ginny Russell of the University of Exeter, this is because it is difficult to tell which toddlers have a very severe disability and which will catch up despite a slow start.

“My recommendation below (two years) is too early to start putting psychiatric labels on a few characteristics like eating behavior,” she said.