• Research Highlight
Autism screening in the first two years of life is an essential tool for ensuring that children and families have access to appropriate supports and services as early as possible. Although effective screening tools are available, researchers are exploring new approaches that could help make early screening more accurate and objective. Research supported by the National Institute of Mental Health suggests that children’s health records may yield some promising insights.
The research team, led by Matthew M. Engelhard, M.D., Ph.D., and Geraldine Dawson, Ph.D., of Duke University, noted that infants’ health care records include health indicators (such as low birth weight) and behavioral indicators (such as challenges with sleeping and feeding) that are often observed in children who later receive an autism diagnosis.
Dawson and colleagues hypothesized that they might be able to use machine learning to incorporate a range of health records measures and develop a predictive model to identify infants who are likely to later receive an autism diagnosis. They further hypothesized that this kind of model might be able to identify children with autism in the first year of life, before standard early autism screening tools can be used.
The researchers analyzed more than 14 years of health records data from the Duke University Health System to develop and evaluate predictive models for early autism detection. The models included a range of possible predictors, including medical visit details, vital signs, procedure codes, and laboratory measurements. For each child, the researchers examined health information recorded at several points in the first year of life: age 30 days, 60 days, 90 days, 180 days, 270 days, and 360 days.
Using documented medical diagnostic codes, the research team identified children who were later diagnosed with autism spectrum disorder, attention-deficit/hyperactivity disorder (ADHD), or other neurodevelopmental conditions.
The study sample comprised a total of 45,080 children, including 924 children with a diagnosis of autism, 10,782 children diagnosed with ADHD or other neurodevelopmental disorders, and a comparison group of 33,374 children who did not meet the criteria for any developmental disorder.
Engelhard, Dawson, and colleagues randomly divided the sample into two subsets, using data from one subset to develop the predictive models and data from the other subset to test the performance of those models. They evaluated performance by comparing statistical model predictions (based on data available in the first year of life) with diagnoses made later in life (based on diagnostic codes).
Using health record data from the first 30 days of life, the model correctly identified about 46% of the infants who were later diagnosed with autism while also correctly identifying about 90% of the infants who did not subsequently receive an autism diagnosis. Using data from the first 360 days of life, the model correctly identified about 60% of children diagnosed with autism (as classified by diagnostic codes) while correctly identifying about 82% of the infants who did not receive a diagnosis.
According to the researchers, the results show that predictive models based on health record data can provide clinically meaningful information earlier than standard early autism screening tools. The researchers also note that their models performed well across the diverse sample. The models correctly identified children diagnosed with autism across races and ethnicities. In addition, the models correctly identified children diagnosed with autism and co-occurring ADHD. This is notable because ADHD and autism have some overlapping features, which can make precise identification more challenging.
Additional research examining how these kinds of models compare to standard early autism screening tools will help clarify whether the approaches identify similar or distinct groups of children on the autism spectrum.
The researchers are continuing to refine these early detection models. Their long-term goal is to develop an objective way to alert health care providers about patients who have a higher likelihood of receiving an autism diagnosis. Health care providers could then increase monitoring of those patients to ensure that they receive appropriate services as soon as possible. The researchers plan to integrate health records-based models with caregiver surveys and other screening tools in clinical practice, and test how parents and providers perceive health records-based approaches to early identification.
Engelhard, M. M., Henao, R., Berchuck, S. I., Chen, J., Eichner, B., Herkert, D., Kollins, S. H., Olson, A., Perrin, E. M., Rogers, U., Sullivan, C., Zhu, Y., Sapiro, G., & Dawson, G. (2023). Predictive value of early autism detection models based on electronic health record data collected before age 1 year. JAMA Network Open, 6(2), Article e2254303. https://doi.org/10.1001/jamanetworkopen.2022.54303
Leave a Reply