detect the autism spectrum disorders (ASD) during early childhood allows early intervention that favors the development and social integration of those affected, and for this reason the American Academy of Pediatrics recommends universal detection of ASD in children aged 18 to 24 months, although diagnosis can currently be delayed until the 3 or 4 years. Now a new study has shown that a suite of automated tools might be able to predict the autism from the baby’s first month of life.
Researchers at Duke University have developed a algorithm using health data from children under one year of age. Babies who were later diagnosed with autism were much more likely to have been examined by an ophthalmologist or neurologist, to have gastrointestinal problems, or to receive physical therapy, than neurotypical children (whose brains develop and function in a way considered normal). Geraldine Dawson, director of the Duke Center for Autism and Brain Development and author of the study, which has been published in JAMA Network Open.
Dawson has explained that his findings also confirm that Autism doesn’t just affect the brain., but to the whole organism, since it can include digestive, sleep, neurological and vision alterations, among others. “We need to think of autism not just as a behavioral condition, but also as a condition that involves physical health,” he says. “This is one way to take advantage of that information: to do a better job of early detection.”
Artificial intelligence for early diagnosis of autism
The researchers’ goal was to design tools that could detect autism at 30, 60, 90, 180, 270, and 360 days of age, by training and evaluating machine learning models using the electronic medical records of 45,080 children cared for in Duke University Health System when they were less than one month old (between January 2006 and December 2020).
“We need to think of autism not just as a behavioral condition, but also as a condition that involves physical health”
The algorithm they developed using this method was able to predict which of these babies would be diagnosed with autism, and even differentiate them from those who would develop ADHD (attention deficit hyperactivity disorder) or other neurological disorders. To determine the model’s performance, these scientists measured sensitivity, specificity, and positive predictive value (PPV) and compared their predictions with caregiver surveys commonly used in autism screening tests.
They looked especially at how the model worked in groups of children who are often missed by standard screening methods, missing out on early diagnosis and losing the benefits that comes with it. 1.5% of the children included in the study (924) met the criteria for autism. Model performance at 30 days of age reached a sensitivity of 45.5% and a PPV of 23% with a specificity of 90%, which was similar to the performance associated with caregiver surveys collected between 18 and 24 months. .
According to Geraldine Dawson if the findings are confirmed in new studies, the algorithm could be used in conjunction with other screening tools, parental reports and medical observations. Because this tool collects information automatically as the child receives care, it could alert a pediatrician that “based on the child’s usage pattern, she has a increased likelihood of a subsequent diagnosis of autism“, it states. And he concludes that “the goal would be for the pediatrician to monitor this baby more closely.”