AI Rare Disease Diagnosis in Children: 18 Cases at Boston

The best number in AI rare disease diagnosis children stories is usually the smallest one. At Boston Children’s Hospital, researchers took 376 patient genomes that had already been analyzed by human specialists without producing a diagnosis, handed each case to OpenAI’s o3 Deep Research model along with filtered gene lists and the clinicians’ own notes, and came away with 18 new diagnoses. That’s roughly 5%. Catherine Brownstein, scientific director of the genetic investigations arm of the hospital’s Manton Center for Orphan Disease Research, put the context plainly in NBC News reporting on June 18, 2026: “It got almost 5% new diagnoses, which doesn’t sound like a lot, but considering how many times these had already been analyzed, that’s a huge number.”

Key Takeaways

• The model was o3 Deep Research, an OpenAI system designed to read many documents and reason across them. It was not a medical model trained specifically on genetics.
• Inputs: full genome sequences, filtered lists of candidate genes, clinical notes, and symptom descriptions for 376 undiagnosed patients.
• Output: 18 new diagnoses — 10 rare neurodevelopmental diseases, 4 neuromuscular disorders, 2 cases of early childhood psychosis, and 2 patients who had died suddenly.
• Seven of the findings were “rediscoveries”: conditions already identified somewhere in the world but not shared in a way the hospital’s team had seen.
• Every AI output was reviewed by humans before becoming a diagnosis. The workflow is AI-assisted, not AI-decided.

What a rare disease is, and why diagnosis takes years

A rare disease in the United States is legally one affecting fewer than 200,000 people, per the Rare Diseases Act of 2002. There are more than an estimated 10,000 of them, and fewer than 5% have an approved treatment.

That combination produces what families call the diagnostic odyssey. A child has symptoms that don’t match any common condition. Specialists rule things out one at a time. Genome sequencing gets ordered, produces a list of genetic variants, and somebody has to decide which variant, out of thousands that look unusual, actually explains this particular child.

The NBC report describes Kyra Benton, who began toe-walking and struggling to run at age 9, developed severe cardiac complications, and needed a tracheotomy at 13. Fifteen years without an answer. The Boston Children’s work identified her condition as myofibrillar myopathy, a disease of the protein scaffolding inside muscle fibers.

How the AI rare disease diagnosis children workflow actually worked

What the model was. o3 Deep Research is a general-purpose reasoning system built to search, read, and synthesize across many sources. That matters for accuracy: this was not a purpose-built genetics model. It was a reading-and-reasoning model pointed at a genetics problem.

What data it got. For each of the 376 cases: the patient’s genome sequence, a filtered candidate-gene list (the human pipeline’s shortlist of suspicious variants), and the clinicians’ written notes describing symptoms, family history, and test results.

What it computed. The hard part of rare-disease diagnosis is not finding variants; sequencing does that. It’s connecting a specific variant to a specific child’s specific symptom pattern, which requires knowing what has been published about that gene anywhere in the world’s literature. A model that has read enormous amounts of biomedical text and can search more of it is genuinely well-suited to that cross-referencing job. For each candidate gene, the model had to construct an argument: this variant, in this gene, produces this protein defect, which has been reported to cause these symptoms, which match this patient.

What it got right. Eighteen cases where the connection had been missed. Ten of those were rare neurodevelopmental diseases, the category where symptom descriptions are hardest to match to genes because the presentations overlap so much.

What it got wrong, or only half-right. Seven of the eighteen were rediscoveries. The condition had been described before, just not in a place the Boston team had found. That’s a real result for the family, and a weaker result for the claim that AI generated new knowledge; the model was very good at retrieval. And the study was retrospective: these were already-collected cases reviewed after the fact, not patients diagnosed in real time.

What the workflow required. Human review of every output. Clinicians at Boston Children’s had to evaluate each proposed diagnosis against the patient before it became a diagnosis. Nothing was accepted because the model said so.

The analogy: the world’s fastest librarian

Imagine a librarian who has read every medical journal ever published and can recall any page in seconds, but has never met a patient and cannot examine anyone.

You hand this librarian a folder: a child’s full genetic readout, a doctor’s notes, and a list of 40 genes that look suspicious. The librarian’s job is not to decide what’s wrong. It’s to say: “In 2011, a group in Finland described three children with a variant in this exact gene and these exact symptoms. Here’s the paper.”

Then a doctor reads that paper, examines the child, and decides. That’s the Boston Children’s workflow. The AI is a spectacular librarian. The diagnosis is still a clinical act.

Eighteen cases, four categories: what the AI found

Category	Number of cases	What made it hard for humans
Rare neurodevelopmental diseases	10	Symptom patterns overlap heavily across dozens of genes; matching requires enormous literature recall
Neuromuscular disorders	4	Progressive, variable onset; often mistaken for other conditions for years (the myofibrillar myopathy case)
Early childhood psychosis	2	Psychiatric presentations rarely trigger genetic workups in the first place
Sudden unexplained death	2	Retrospective analysis only; no living patient to re-examine
Of these, rediscoveries	7	Condition was known globally but not connected to this case locally

The last row is the honest one. Eleven were genuinely novel connections for the field; seven were connections that existed in the literature and hadn’t reached the right desk.

How to Teach Your Kid About AI in Medical Diagnosis

Ages 5–8: The Matching Mystery

Make 10 cards with simple symptom pictures (rash, cough, limp) and 10 cards with silly “cause” names. Secretly set up correct pairs. Let your kid match them by trial and error, then show them a “book” (a list you wrote) that gives the answers. Ask which was faster, guessing or looking it up. That contrast is the whole story: the AI was a very fast book, not a magic doctor.

Ages 9–12: The Needle in 3,000 Haystacks

Print a page with 200 nonsense “gene names,” one of which appears in a second list you give them of “genes that cause limping.” Time how long it takes to find the overlap by hand. Then point out that real genome analysis compares thousands of variants against millions of published findings. Ask: what would you need to make this fast? (Answer: a system that has already read everything. That’s the model.)

Ages 13+: Retrieval Versus Discovery

Give your teen the numbers: 376 cases, 18 diagnoses, 7 of them rediscoveries. Ask them to separate “the AI found something nobody knew” from “the AI found something somebody knew but this team hadn’t seen.” Then have them read the abstract of the August 2026 scoping review of AI in pediatric rare disease in the Journal of Medical Internet Research and find the validation statistics: only 25.9% of 81 studies included external validation, and just 14.8% included prospective evaluation. Ask why that matters.

The question to ask: “Seven of the 18 answers already existed in a journal somewhere. Does that make the AI more impressive or less?”

What to actually do at home

Understand that your pediatrician may already be using something like this

AI decision-support tools are in clinical workflows now, mostly invisibly. If your child is in a diagnostic process, it’s a reasonable and non-adversarial question to ask: “Are any AI tools involved in reviewing this, and who checks the output?” Good clinicians answer that easily. Our piece on AI in pediatric diagnosis covers what’s already deployed.

Don’t run your child’s symptoms through a chatbot and treat the answer as information

The Boston Children’s result depended on inputs no family has: a sequenced genome, a professionally filtered variant list, and clinical notes. Without those, a chatbot is generating plausible text about symptoms, not performing the task that produced 18 diagnoses. The gap between those two activities is enormous.

Teach the difference between retrieval and reasoning

This case is unusually clean for that lesson. The model’s strongest capability here was finding a published connection. That’s retrieval. Kids who can name that distinction will read the next decade of AI-medicine claims much more accurately. The related idea is covered in our retrieval-augmented generation explainer.

If you’re in a diagnostic odyssey, ask about reanalysis

The most actionable thing in this story for families: genome data can be reanalyzed later, as knowledge grows. All 376 of those cases had been analyzed before. Asking whether and when reanalysis happens is a fair question for a genetics team.

What not to do

Don’t tell your kid “AI diagnosed 18 children” without the second half of the sentence. It didn’t diagnose anyone; it proposed 18 candidate answers that human clinicians then evaluated and confirmed. In pediatric medicine specifically, that distinction is not pedantic. It’s the entire safety model.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can explain that the AI was given a genome plus doctors’ notes, and that a doctor confirmed every answer.
• Month 2 red flags: They start treating a chatbot’s health answers as diagnostic. Point back to the inputs Boston Children’s had and the ones a phone doesn’t.
• Month 3 self-check: Ask them what would have to be true for a study like this to count as clinical proof. If they mention testing on new patients rather than old files, they’ve understood prospective validation, which is the single biggest gap the JMIR review identified.

Frequently Asked Questions

Did the AI diagnose children by itself?

No. The model produced candidate diagnoses from genome data, filtered gene lists, and clinical notes. Clinicians at Boston Children’s reviewed every output and confirmed each diagnosis against the patient. The published framing is AI-assisted diagnosis, and human review was a required step, not a courtesy.

Why is 5% considered a good result?

Because all 376 cases had already been analyzed by human genetics experts, sometimes repeatedly, without a diagnosis. The baseline wasn’t “no one looked”; it was “specialists looked and came up empty.” Brownstein’s point was that finding 18 answers in an already-picked-over set is unusual.

What’s the difference between this and DeepRare?

DeepRare is a separate agentic system reported in Nature in February 2026 that coordinates about 40 specialized tools. On 163 difficult cases it reached 64.4% first-attempt accuracy versus 54.6% for five physicians with 10-plus years of experience, and rare-disease specialists endorsed its reasoning 95.4% of the time. Boston Children’s used a general reasoning model on already-failed cases; DeepRare is a purpose-built diagnostic system benchmarked head to head.

Should families push for AI review of an undiagnosed case?

The reasonable version of that request is to ask about periodic reanalysis of existing genome data, which is standard practice at many centers regardless of AI. Asking specifically for “AI review” is less useful than asking when the data will be looked at again and against what updated knowledge.

How reliable is AI for rare disease diagnosis overall?

Less than headlines suggest. The August 2026 JMIR scoping review of 81 studies found 93.8% reported internal validation but only 25.9% included external validation and 14.8% prospective evaluation. For differential diagnosis in rare pediatric case reports, large language models hit top-1 accuracy of just 8.2% to 13.1%. The authors conclude these results show “technical promise rather than established clinical effectiveness.”

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. NBC News. (2026, June 18). “AI helped Boston Children’s Hospital diagnose rare diseases in kids.” https://www.nbcnews.com/tech/innovation/ai-boston-childrens-hospital-diagnose-rare-diseases-kids-openai-rcna350387
2. Zhao, J., Luo, J., Li, Q., & Chen, Y. (2026, August 28). “Machine Learning, Large Language Models, and Multimodal AI for Diagnosing Pediatric Rare Diseases: Scoping Review.” Journal of Medical Internet Research. https://pmc.ncbi.nlm.nih.gov/articles/PMC13524366/
3. National Institutes of Health, Genetic and Rare Diseases Information Center. “About GARD.” https://rarediseases.info.nih.gov/about
4. DeepRare research team. (2026). Agentic diagnostic framework for rare disease. Nature. DOI 10.1038/s41586-025-10097-9. Coverage: https://medicalxpress.com/news/2026-02-deeprare-ai-outperforms-doctors-rare.html
5. Canavati, C., Tabach, Y., et al. (2026). EvORanker: evolutionary-signal ranking of candidate disease genes. Genetics in Medicine. Coverage: https://www.bfhu.org/2026/03/30/a-new-ai-breakthrough-could-change-the-long-search-for-rare-disease-diagnoses
6. MedlinePlus Genetics (NIH). “Myofibrillar myopathy.” https://medlineplus.gov/genetics/condition/myofibrillar-myopathy/
7. Rare Diseases Act of 2002 (definition of rare disease in the United States). https://rarediseases.info.nih.gov/about