Autism spectrum disorder (ASD) is the most prevalent childhood developmental disorder in the United States. For decades, there were many individuals with autism who were never diagnosed or treated for the condition. Because of increased awareness and screening tools, the number of children who are diagnosed with autism by age 8 has risen from 1 out of 150 in the year 2000 to 1 out of 54 in 2016, according to the Centers for Disease Control and Prevention (CDC).
The age that experts can reliably diagnose autism has gone down dramatically during this period as well, says David S. Mandell, a doctor of science and a professor of psychiatry at the Perelman School of Medicine at Penn Medicine and the associate director of the Center for Autism Research at the Children’s Hospital of Philadelphia. According to the Southwest Autism Research & Resource Center, the average age of diagnosis has decreased significantly from 1997 to 2017, from 4 years and 4 months to 3 years and 10 months.
“The excitement about earlier diagnosis is linked to the importance of early intervention. In some cases, signs of autism can emerge at 9 months of age or even earlier,” says Dr. Mandell.
Infants and toddlers learn at a tremendously fast rate, and if autism isn’t diagnosed early, it’s harder to go back and teach language and skills that were missed, notes Mandell. “If you think about the way we absorb material, if you learn something at a regular pace, you’re likely to understand and retain it much better than if you cram for it at the end." The same goes for children with autism, he adds.
The earlier the intervention, the better the outcome. “If we intervene early, we can change the trajectory of disease, leading to improved outcomes and better and happier lives,” Mandell explains.
RELATED: Autism Diagnosis: What to Expect When It Comes to Screening and Testing
Machine Learning: What It Is and How It Helps in Diagnosing Autism
Much of the new research around autism diagnosis uses something called machine learning, a broad set of statistical approaches that can be used in many fields. In medicine, machine learning can be used to classify and predict diagnoses and outcomes for different conditions, from distinguishing different stages of brain cancer to improving cardiovascular risk prediction, according to Suzanne Macari, PhD, a research scientist and the codirector of the Yale Social and Affective Neuroscience of Autism Program at the Child Study Center in New Haven, Connecticut.
“Supervised machine learning uses algorithms that can learn to uncover patterns in large data sets and make predictions about a set of given outcomes,” Dr. Macari explains. This method holds promise for autism research, especially today in the era of “big data,” when large data sets are frequently shared across institutions, she adds.
High Risk Infants Provide Clues for Earlier Autism Diagnosis
Macari’s research has utilized machine learning in three different studies to explore autism symptoms, diagnosis, and the different ways that the syndrome develops in infants who are at high risk for autism because they have an older sibling with ASD.
“The risk of having autism is about 20 percent if you have a sibling diagnosed with autism,” says Antonio Hardan, MD, a professor of psychiatry and behavioral sciences at Stanford University Medical Center in Stanford, California, and the director of the autism and developmental disabilities clinic. For the general population, the risk is 1.5 percent, or about 1 in every 68 individuals, according to the CDC.
“In our first study, we set out to predict high levels of ASD symptoms at 24 months from infants’ behavior during a social-communication evaluation at 12 months,” says Macari. The aim of the research, published in Journal of Autism Developmental Disorders, was to help identify infants who were most likely to need intervention in the second year and to identify early behaviors that suggest a baby will most likely develop typically, she says.
“It turned out that identifying infants who would become symptomatic of autism by age 2 is complicated because the emergence of symptoms is so variable across children,” says Macari. They did find small sets of social-communication behaviors at 12 months that all predicted an autism classification at 24 months.
Infants who did not have any clear instances of showing objects to others during a play session or participating in a moment of shared attention on an object were at an increased risk for developmental delays and ASD. “In addition, we pinpointed specific behaviors that, if present robustly in a child, appeared to herald a typical developmental trajectory,” she says. These behaviors were: engaging in at least one instance of showing objects (toys) to others during a play session; initiating shared attention of an object (looking at the object, then looking to an adult, then back at the object again); and having an age-appropriate ability to regulate attention and activity level.
In a second study, parents were asked to fill out a questionnaire when the higher-risk infant was 12 months old to see if there were behaviors that might predict a later autism diagnosis. The study, published in the Journal of Autism Developmental Disorders in January 2015, identified a small number of parent-reported behaviors that would merit more significant concern about future ASD symptoms, says Macari. When parents reported a decline in play, communication, and impaired vocal imitation, it correctly classified a majority of ASD cases with high specificity, according to the study authors.
Most recently, machine learning was used in large study of high-risk 18-month-old infants from across the United States and Canada from the Baby Siblings Research Consortium (BSRC). Investigators found that there is no “one way” that autism presents, but instead identified three distinct combinations of features at 18 months that were predictive of an ASD outcome: poor eye contact combined with lack of communicative gestures and giving; poor eye contact combined with a lack of imaginative play; and lack of giving and presence of repetitive behaviors, but with intact eye contact. The research was published in December 2015 in the Journal of the American Academy of Child and Adolescent Psychiatry.
“A major implication of our research is that we should abandon the approach of using a single set of behaviors to identify all infants developing ASD. In other words, screening tools must allow different combinations of features to indicate risk for the disorder,” says Macari.
RELATED: What Causes Autism? The Possible Genetic and Environmental Risk Factors
Brain Imaging of Higher-Risk Sibling Infants
New research is also examining behavior patterns along with brain imaging for clues on how the condition manifests in brain activity in the younger siblings of children with autism.
In a study known as the Infant Brain Imaging Study (IBIS), investigators assess the babies at 6 months, 12 months, and 24 months, says Dr. Hardan, who is not involved in this research. “They have MRIs performed at each age, as well evaluation in order to identify behavior, cognitive, and imagining markers of infants who later on develop autism,” he says. “By carefully monitoring the development of the siblings after their birth, it will allow us to identify potential markers of autism later on in life."
Interventions for Higher-Risk Siblings Before Autism Diagnosis
New research is also exploring if autism intervention and therapies may improve the outcome for infants before they are even diagnosed with autism, with the hopes that early intervention might change the trajectory of these infants, says Hardan.
“For example, if you need five symptoms to make a diagnosis of autism, sometimes people are starting interventions when there are only two or three symptoms in the hopes to change the trajectory of the disorder and optimize the long-term outcomes,” he says.
A small study published in the Journal of Autism Development Disorders found that early intervention in infants between 6 and 15 months old who had at least two symptoms along with clinical concerns of autism was associated with improvements in symptoms and language development rates when compared with infants who had no intervention.
RELATED: Study Finds That Routine Pediatrician-Administered Screenings Could Flag Autism Earlier Than Other Methods
The Future of Autism Screening and Diagnosis
The gold standard method to diagnose autism remains a direct evaluation by experienced clinicians, according to Macari. “There is simply no substitute for this, as a clinical best estimate (CBE) diagnosis involves the integration of several sources of information about a child, including complex factors such as developmental profile, adaptive skills, medical history, and comorbid symptoms,” she says.
Some recent studies have explored the idea of using machine learning to refine existing evaluation tools or aid in the development of new tools in order to gear the instruments to children of different ages, gender, developmental level, and language level, says Macari. “This is a particularly exciting direction, as it aligns with ongoing work in our group,” she adds.
Future ways to screen for autism will have content adjusted to the characteristics of each child, such as age, language level, or other features, as well as screeners that account for different presentations of autism, according to Macari. “All these advancements are likely to depend heavily on the use of machine learning, data mining, and other data science approaches,” she says.