Medtech’s AI Revolution is Leaving Kids Behind

“A person’s a person, no matter how small.”—Dr. Seuss

There was clearly something different about the boy in Colleen Kraft’s exam room.

He was there with his younger sister, an unusually quiet toddler. Both were new patients, having moved to southern California from another state.

The siblings suffered from behavioral and communication issues and ended up in Kraft’s office—by chance or by choice—seeking professional diagnoses for insurance and treatment purposes. The meeting proved fateful for the boy and his sister: Their new medical champion was a 35-year-plus practicing pediatrician and noted expert in early brain/childhood develop-ment and behavioral health.

Kraft also was (is) a medical technology aficionado, forever watchful for the innovations that improve patient care. She is particularly interested in solutions that improve screening for pediatric patients with behavioral or developmental issues.

“If there are some technologies that can really help out…in terms of socialization, in terms of reading, in terms of other types of medical diagnoses they have—great,” Kraft, M.D., an American Academy of Pediatrics past president and professor of pediatrics at USC’s Keck School of Medicine, told AMA News Wire in December. “Technology that helps me get the right children in the right lines when it comes to specialists is going to be really important, and anything that we can do diagnostically—or better screening within primary care—that will be efficient and cost-effective and take the burden off our patients.”

Kraft employed such technology to diagnose autism in the four-year-old boy and speech apraxia in his sister, and direct the pair to the necessary treatments. Kraft procured both prognoses through an artificial intelligence- and machine learning-driven diagnostic aid for autism spectrum disorder (ASD), a neurological and developmental condition that affects communication, learning, and behavior. Although autism can be diagnosed at any age, it is considered a “developmental disorder” because symptoms generally appear in the first two years of life, according to the National Institute of Mental Health.

Developed by Palo Alto, Calif.-based Cognoa Inc., Canvas Dx uses quantitative social, behavioral, linguistic, and executive functioning data from both caregivers and clinicians to diagnose ASD. The tool, according to the company, demonstrates equitable performance across race, ethnicity, socioeconomic status, and geography, thus tackling existing biases prevalent in conventional autism diagnoses. Canvas Dx boasts a 95% negative predictive value and 94% positive predictive value, and guards against false classifications by abstaining from a predictive diagnosis in the absence of sufficient data. 

Earning U.S. Food and Drug Administration (FDA) Breakthrough Device Designation in 2019 and De Novo classification two years later, Canvas Dx is touted as the only device authorized by the agency as an ASD diagnostic for children ages 18 months through 5 years.

It is also one of the few AI-enabled devices authorized for children. Not surprising, considering pediatric-specific medical products (and manufacturers) are a rarity in the medtech world.

Despite the proliferation of medical products incorporating artificial intelligence, only a tiny fraction are specifically designed for kids; moreover, a JAMA Pediatrics study found that only one in five devices approved for children actually incorporated pediatric data into their algorithms.

“Despite rapid growth in the availability of AI/ML-enabled devices in recent years, only a small number have been authorized for pediatric use,” the study states. “Among devices labeled for pediatric patients, few device manufacturers disclosed information in regulatory documents on whether algorithm validation was performed in pediatric cohorts…” 

Researchers from the Boston Children’s Computational Health Informatics Program analyzed 876 AI/ML-enabled medical devices identified on the FDA’s website as of March 2024. Of that total, 549 devices (62.7%) gained authorization after 2021, and 853 (97.4%) earned 510(k) clearance. 

One-third of device authorizations (292, or 33.3%) were specifically approved for adults, and nearly half (435) made no mention of pediatric use. Most products focused on radiology (79.9%), neurology (8.7%), and cardiovascular (6.7%) specialties.

The study identified 149 AI/ML-enabled medical devices labeled for pediatric use, or 17% of all AI/ML-enabled device authorizations. Overall, nearly 19% of those products explicitly reported development using pediatric patient datasets, and 14.8% used only adult data during development. The remaining 66.4% of devices did not report whether pediatric patients were studied. 

 Approval summaries described patients’ ages in the datasets for 54 pediatric devices (36.2%). 

The study warns that such significant disparities between adult and pediatric AI-enabled medtech products may endanger juvenile patients. “The current regulatory framework may expose children to off-label use, differential performance of algorithms, and safety risks,” the authors conclude. “Additionally, the lack of standardized reporting of pediatric device characteristics precludes informed decision-making by healthcare clinicians on appropriate device use.”

Children have long been medtech’s forgotten folk. Now they are being left out of the AI revolution, too.

Artificial intelligence and machine learning promise to transform healthcare delivery, but there are still limits to its capabilities. For now.