Tools & Strategies News

Predictive Analytics Tools Accurately Detect Pediatric Autism

A set of automated tools can predict early autism using EHR data collected before a patient’s first birthday, according to new research.

predictive analytics pediatric autism diagnosis

Source: Getty Images

By Shania Kennedy

- A study published last week in JAMA Network Open describes how a set of EHR data-based predictive analytics tools can detect early autism using patient data collected before 1 year of age.

According to the study, early detection of autism in childhood is crucial for ensuring patients and their families have access to appropriate behavioral support and resources, which are associated with improved outcomes. To improve early detection rates, the American Academy of Pediatrics recommends universal screening among infants from 18 to 24 months.

Despite this recommendation, the Centers for Disease Control and Prevention (CDC) report that in 2018, the median age at earliest known autism diagnosis was 50 months, which the authors posit indicates that most children with autism are being identified too late to benefit fully from early support interventions and resources.

Standard early autism screening tools are effective, but many are recommended at 16 to 30 months of age, even though some children would benefit from earlier diagnosis, the researchers noted. These tools can also contribute to disparities in diagnosis by performing worse among certain populations, such as girls, racial and ethnic minority children, and children from lower-income households,

To combat this, the researchers set out to develop tools designed to detect autism at 30, 60, 90, 180, 270, and 360 days of age using routinely collected EHR data. To develop these tools, they gathered EHR data from 45,080 children seen within the Duke University Health System before 30 days of age between January 2006 and December 2020.

To evaluate model performance, the research team measured sensitivity, specificity, and positive predictive value (PPV). Model predictions were compared with caregiver surveys typically used in autism screenings.

Of the 45,080 children included in the study, 924, or 1.5 percent, met autism criteria. Model performance at age 30 days achieved 45.5 percent sensitivity and 23 percent PPV at 90 percent specificity, which was similar to the performance associated with caregiver surveys collected at 18 to 24 months.

The model’s accuracy improved significantly when evaluating 1-year-old patients at different statistical specificities, or the ability of a test to correctly identify people without a condition or disease, achieving 59.8 percent sensitivity and 17.6 percent PPV at 81.5 percent specificity and 38.8 percent sensitivity and 31 percent PPV at 94.3 percent specificity.

These findings suggest that EHR-based autism detection could be integrated with caregiver surveys to improve the accuracy of early autism screening and that early correlates of autism found in routinely collected EHR data could aid in passive, predictive model-based monitoring to improve the accuracy of early childhood autism detection, according to the researchers.

This research contributes to a growing number of studies leveraging artificial intelligence (AI) and predictive analytics to support autism diagnosis.

In May, a study published in npj Digital Medicine found that an AI-based software-as-a-medical-device (SaMD) could assist clinicians in primary care settings accurately diagnose autism spectrum disorder (ASD) in children up to 6 years old.

The tool uses behavioral features predictive of ASD, sourced from a caregiver questionnaire, a healthcare provider questionnaire, and two short home videos, to offer recommendations for the primary care provider.

The tool classifies patients as ASD positive, ASD negative, or “indeterminate,” which indicates that the information input was insufficient for the algorithm to render a highly predictive output.

Of the cohort who received an ASD diagnosis by a specialist, 52.5 also received a determinate result from the tool. The device correctly classified all the cases except for one false negative. Of those participants who got an ASD negative and neurotypical diagnosis by the specialist, 35 percent received an ASD negative result by the tool, and none were misclassified as ASD positive.

The researchers concluded that AI-based diagnostic aids may have significant potential to assist clinicians in primary care settings with ASD diagnosis, but further research is needed.