A June 18, 2026 peer-reviewed study published in NEJM AI details how an OpenAI o3 Deep Research reasoning model helped clinicians diagnose 18 previously unsolved rare childhood genetic diseases out of a 376-case cohort of long-unsolved cases. The work adds a 4.8% absolute diagnostic yield atop years of prior specialist analysis for cases in the study cohort, a result that is statistically significant with a p-value of less than 0.05 OpenAI’s rare childhood disease diagnosis study announcement.
The study was a collaboration between the Manton Center for Orphan Disease Research at Boston Children’s Hospital, Harvard Medical School, and OpenAI. Researchers used de-identified, HIPAA-compliant clinical and genomic data from the 376 long-unsolved cases included in the Manton Center’s long-term case registry, all of which had undergone years of prior specialist analysis without a confirmed diagnosis OpenAI’s rare childhood disease diagnosis study announcement.
How does AI help physicians diagnose rare genetic diseases affecting children?
To run the analysis, the research team assembled a standardized data packet for each of the 376 long-unsolved cases in the Manton Center’s registry. Each packet included Human Phenotype Ontology (HPO) terms to codify the patient’s clinical presentation, relevant clinician notes, and demographic metadata OpenAI’s rare childhood disease diagnosis study announcement.
Every variant in the filtered table was limited to a minor allele frequency (MAF) below 0.1% in the gnomAD population database. Each variant entry included predicted protein effect, ClinVar classification, and segregation signal quality across available family members OpenAI’s rare childhood disease diagnosis study announcement.
All candidate findings were reviewed by at least two board-certified clinical geneticists on the study team, using the ACMG/AMP (American College of Medical Genetics and Genomics/Association for Molecular Pathology) variant classification framework, the standard for clinical genetics laboratories. Disagreements between reviewers were resolved by consensus, and no model output was ever classified as a confirmed diagnosis without further validation OpenAI’s rare childhood disease diagnosis study announcement.
A result was only counted as a formal diagnosis after expert review classified the variant as pathogenic or likely pathogenic per ACMG/AMP standards. It also required confirmation by a CLIA-certified clinical laboratory and formal return of the result to the patient’s family, per the study’s guardrails OpenAI’s rare childhood disease diagnosis study announcement.
What was the diagnostic yield of the AI workflow for long-unsolved cases?
The research team deployed the validated workflow across four pre-specified patient groups drawn from the total 376-case cohort. Specifically, these groups included pediatric patients with neurodevelopmental disorders, individuals with rare neuromuscular conditions, children and adolescents with early-stage psychosis, and cases of sudden unexplained death in pediatric populations OpenAI’s rare childhood disease diagnosis study announcement.
The 18 new diagnoses generated by the workflow represent a statistically significant 4.8% absolute additional diagnostic yield over earlier specialist review, with a p-value of less than 0.05. This translates to 1 new confirmed diagnosis for every 21 families in the study cohort who had previously received no answers for their child’s condition OpenAI’s rare childhood disease diagnosis study announcement.
How does this study align with OpenAI’s broader health AI safety work?
The study builds on OpenAI’s broader health AI roadmap, which includes recent improvements to health intelligence in ChatGPT rolled out to Enterprise and Team tiers in April 2026. Those improvements cut flagged factuality issues in health responses by 71% over the subsequent two months of production traffic, per the company OpenAI’s health intelligence improvement announcement.
All candidate findings from the rare disease workflow still require dual expert review per ACMG/AMP standards. They also require CLIA-certified lab confirmation before being returned to patients, as the model is designed only to support, not replace, clinical geneticists OpenAI’s rare childhood disease diagnosis study announcement.
Bottom line: Clinical genetics programs can pilot this guardrailed OpenAI o3 Deep Research workflow to reanalyze long-unsolved pediatric rare disease genomic cases, adding a 4.8% absolute diagnostic yield for families who previously received no answers for their child’s condition. All candidate findings require dual board-certified clinical geneticist review per ACMG/AMP standards and CLIA-certified lab confirmation before return to patients, as the model is designed only to support, not replace, clinical expert analysis.
