Researchers have developed a screening system that they say is about 80% accurate in children under the age of two, and say artificial intelligence could help experts identify young children who may have autism.
The researchers say their approach, which is based on a type of AI called machine learning, could pay off.
“Using AI models, we can leverage available information to identify individuals at high risk of autism earlier, enabling them to receive earlier diagnosis and support,” said Dr Christina Tamimis, study co-author from the Karolinska Institutet in Sweden.
But, she added, “we want to stress that this algorithm cannot diagnose autism, as this[still]should be done with gold-standard clinical methods.”
This isn't the first time researchers have tried to use AI to screen for autism: Scientists have previously combined the technology with retinal scans of children.
Tamees and his colleagues report in the journal Jama Network Open how they used data from the Spark study, a US research initiative that includes information on 15,330 children with autism diagnosis and 15,330 without.
The researchers describe their method, which focuses on 28 indicators that can be easily obtained before a child turns 24 months of age, based on parent-reported information from medical and background questionnaires, such as the age at which the child first laughed.
The team then created a machine learning model that looked for different patterns in combinations of these features between autistic and non-autistic children.
The team used the data to build, tune, and test four different models, then selected the most promising one and tested it on a dataset of 11,936 participants for whom data on the same features was available. A total of 10,476 of these participants had been diagnosed with autism.
Results showed that overall, the model correctly identified 9,417 (78.9%) participants with and without autism spectrum disorder, with accuracy rates of 78.5% for children up to 2 years old, 84.2% for 2- to 4-year-olds, and 79.2% for 4- to 10-year-olds.
Further testing using a different dataset containing 2,854 autistic individuals revealed that the model correctly identified autism diagnoses 68% of the time.
Tamimees said: “This dataset was a different research cohort of families with just one child with autism, and we found that it was missing some parameters, which led to a slightly lower performance and required further development.”
The researchers said that indicators that generally seemed most important in terms of the model's predictions included problems with eating, age at which a child first constructed a long sentence, age at which a child achieved toilet training and age at which a child first smiled.
The team added that additional analyses comparing participants whom the model correctly identified as having autism with those whom it incorrectly identified as not having autism suggested that the model tended to identify individuals with more severe symptoms or more common developmental problems as having autism.
But some experts urged caution, pointing out that the model is only 80% good at identifying non-autistic people, meaning 20% may be falsely identified as possibly autistic, and that pushing for early diagnosis could be problematic.
Prof Ginny Russell, from the University of Exeter, said this was because it was difficult to determine which young children were seriously impaired and who would “catch up” despite a slow start.
“My recommendation is that before age 2 it's too early to assign a psychiatric label based on some indicators like eating behavior,” she said.
