Your Pediatric AI Tool Was Trained on Adults

A three-year-old with a fever of 39.2 and a heart rate of 155 walks into your emergency department. The AI-powered clinical decision support tool on your screen flags the heart rate as elevated but returns a low acuity score. The algorithm does not know that a heart rate of 155 is within the expected range for a febrile toddler. It was trained on adults, where 155 means something very different. This is not a theoretical concern. It is happening now, in departments that have adopted AI tools without asking the most basic question: was this built for my patient population?

The Adult Data Problem

The vast majority of AI diagnostic and clinical decision support tools entering the ED were developed using adult patient datasets. This is not a minor technical footnote -- it is a foundational flaw when those tools are applied to children. Pediatric physiology is not simply a scaled-down version of adult physiology. Vital sign norms shift dramatically with age. A blood pressure of 80/50 may be perfectly normal in a two-year-old and represent decompensated shock in a forty-year-old. Respiratory rates, heart rates, and even laboratory reference ranges are age-dependent in ways that most AI training datasets do not account for.

A 2024 scoping review of pediatric sepsis prediction technologies found that 85% of the 27 studies examined were single-center investigations, and 59% relied on logistic regression -- a method with known limitations when applied to the complex, nonlinear trajectories of pediatric disease. Seventy-four percent of the studies used datasets with a low prevalence of sepsis-related outcomes, with AUROC scores ranging wildly from 0.56 to 0.99. That range tells you something important: the field does not yet have reliable, validated tools for predicting pediatric sepsis with AI.

Research published in Surgery Open Science demonstrated that pediatric physiological data requires age-normalization before AI models can perform comparably to their adult counterparts. Without that normalization, the models produce unreliable outputs. The researchers developed a Continuous Age-Normalized SOFA score to bridge the gap, achieving an accuracy of 0.84 on pediatric populations -- but this kind of preprocessing is the exception, not the standard. Most commercially available AI tools do not perform it.

What Pediatric AI Gets Wrong

The failure modes are predictable if you understand pediatric emergency medicine. Children do not present the way adults do, and the differences are not subtle. A toddler with appendicitis may present with nothing more than irritability and refusal to walk. A neonate in early sepsis may be hypothermic rather than febrile. An adolescent with diabetic ketoacidosis may present with abdominal pain that mimics a surgical abdomen. These atypical presentations are the rule in pediatrics, not the exception.

AI tools trained on adult presentations lack the contextual framework to interpret these patterns. A study in the American Academy of Pediatrics' Hospital Pediatrics journal found that while clinicians recognized the potential of AI-CDS tools to improve diagnostic accuracy and standardize care, they raised significant concerns about the ability of AI-based tools to appreciate nuanced pediatric care and accurately interpret pediatric data. The tension between algorithmic recommendations and clinician autonomy was a recurring theme -- physicians worried that less experienced clinicians might defer to a confident-sounding AI output that is, in fact, wrong for the pediatric patient in front of them.

Developmental variation compounds the problem. AI models must account for the physiologic changes and shifts in disease risk that occur from infancy through adolescence. Automated interpretation of pediatric radiographs must contend with maturational changes in skeletal anatomy, differing appearances of common pathologies, and broader differential diagnoses than adult imaging presents. A growth plate that looks like a fracture to an adult-trained algorithm is not a rare edge case -- it is a daily occurrence in a pediatric ED.

The Governance Gap

An analysis of FDA-cleared pediatric software as a medical device shows a striking imbalance: radiology submissions dominate at 92.4%, while other specialties -- the ones where AI could arguably provide the most clinical value in the ED -- lag significantly behind. Cardiovascular devices focused primarily on monitoring (70%), with diagnostic applications less prominent. Neurology devices showed a strong diagnostic focus (63.3%) but still had meaningful monitoring components. The critical specialties of emergency medicine and pediatric acute care are largely absent from the FDA-cleared landscape.

This governance gap matters because the consequences of AI errors in pediatrics are disproportionately severe. A missed diagnosis in a child does not just affect the current episode -- it can reduce decades of quality-adjusted life years. Off-label use of AI tools is common in pediatrics, mirroring the off-label medication patterns that already characterize the specialty. When slight distributional shifts in data can substantially impact predictions, applying adult-validated AI to pediatric populations without dedicated validation is a form of off-label use that carries real risk.

The Journal of Medical Internet Research reported that emergency physicians are already engaging with commercially available generative AI models to bolster diagnostic confidence, including for rare or infrequently encountered diagnoses. But as one emergency physician noted, "Most of these commercially available tools are being trained on all the data that's on the internet...you run the risk of it generating things based on opinions people have shared on Reddit or Twitter." Applying that uncertainty to pediatric patients -- who cannot advocate for themselves and whose presentations are inherently more ambiguous -- is a risk profile that warrants deliberate governance, not ad hoc adoption.

What You Should Be Doing Now

Before you allow any AI clinical decision support tool to influence care for pediatric patients in your department, ask the vendor one question: what percentage of your training data came from pediatric populations, and what age ranges were represented? If they cannot answer with specificity, or if the answer is less than 10%, that tool has not been validated for the patients you are treating. Treat it accordingly -- as an unvalidated instrument that requires physician override for every pediatric case.

Build age-stratified review into your quality assurance process. When AI-assisted decisions result in unexpected outcomes for pediatric patients, do not simply log them as general QA events. Track them separately. Look for patterns by age group. A tool that performs well for adolescents may fail catastrophically for neonates, and your QA data will not reveal that unless you are stratifying.

Educate your team -- residents especially -- about the specific failure modes of AI in pediatric presentations. The most dangerous scenario is not an AI tool that is obviously wrong. It is one that is confidently wrong in a way that aligns with what an inexperienced clinician might already believe. A resident who sees an AI output confirming their suspicion that a febrile infant is low-acuity may not apply the clinical scrutiny that the presentation demands. Make AI skepticism part of your pediatric teaching.

Advocate within your institution for pediatric-specific validation before AI tool deployment. The governance frameworks exist. The pediatric-centric approaches emphasizing transparency, inclusive participation, equitable data practices, and rigorous post-deployment monitoring have been articulated. What is missing is institutional will to apply them before, rather than after, something goes wrong.

Dr. Chet's Take

I have spent my career in environments where the margin for error is measured in minutes and the patients cannot always tell you what is wrong. In air medical transport, in telemedicine network oversight, in forward-deployed military medical operations -- the common thread is that you build systems for the population you are actually serving, not the population that is easiest to study. Pediatric emergency medicine demands the same discipline. When I run a clinical operation, I do not accept equipment that has not been tested for the conditions we will face. AI tools are no different. If a vendor tells me their product works in the ED but cannot show me pediatric validation data, that product is not ready for my department. The stakes are too high and the patients are too vulnerable for anything less than that standard.

AI Won't Wait. Neither Should You.

The integration of AI into emergency medicine is accelerating, and the gap between what these tools promise and what they deliver for pediatric patients is widening. If you lead an emergency department, direct a residency program, or treat children in any acute care setting, understanding these limitations is not optional -- it is a core competency. Consider enrolling in my course: AI in Emergency Medicine: Becoming AI Bulletproof.

Learn more: AI in Emergency Medicine: Becoming AI Bulletproof

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice -- not just the hype -- I put together a free practical guide. You can download it here: AI in EM Survival Guide.

Dr. Chester "Chet" Shermer, MD, FACEP is a Professor of Emergency Medicine, TeleHealth, HEMS and Critical Care Transport, and State Surgeon for the Army National Guard. He is the founder of Global MedOps Command and the creator of AI in Emergency Medicine: Becoming AI Bulletproof.

Sources:

1. Pediatric Research, "AI-based clinical decision support in pediatrics," https://pmc.ncbi.nlm.nih.gov/articles/PMC9668209/
2. Surgery Open Science, "Children are small adults (when properly normalized)," https://pmc.ncbi.nlm.nih.gov/articles/PMC10561114/
3. Nature, "Scoping review on pediatric sepsis prediction technologies," https://www.nature.com/articles/s41746-024-01361-9
4. Nature, "Toward governance of AI in pediatric healthcare," https://www.nature.com/articles/s41746-025-02000-7
5. AAP Hospital Pediatrics, "Clinician Perspectives on Decision Support and AI-based Decision Making," https://publications.aap.org/hospitalpediatrics/article/14/10/828/199452/
6. JMIR, "The Potential and Peril of AI in the Emergency Department," https://www.jmir.org/2025/1/e89200