A study shows an additional 4.8% diagnostic yield in previously unsolved cases using an artificial intelligence-assisted research workflow.
Researchers from Boston Children’s Hospital, Harvard University, and OpenAI used an artificial intelligence (AI) reasoning model to reanalyze 376 previously unsolved genetic cases, identifying leads that resulted in 18 new diagnoses.
The study, published in NEJM AI, utilized the OpenAI o3 Deep Research model to analyze de-identified clinical and genomic data. The AI-assisted workflow provided a 4.8% diagnostic yield after the cases had already undergone analysis by specialists.
The research team designed the workflow to act as an explanation-first reasoning layer on top of existing genomic pipelines. Instead of returning only a ranked gene, the model connected clinical features, inheritance patterns, variant evidence, and scientific literature to create justifications for human reviewers to evaluate.
“The bottleneck is time. An expert can devote only so much of their day to any one particular person,” says Catherine Brownstein, PhD, Boston Children’s Hospital’s Manton Center for Orphan Disease Research, in a release.
Workflow and Laboratory Confirmation
For each case, the team assembled a de-identified packet containing standardized Human Phenotype Ontology terms to describe clinical presentations, clinician notes, and a filtered variant table. The table captured variant rarity, predicted effects on encoded proteins, and ClinVar classification.
The researchers reviewed the model’s outputs using the American College of Medical Genetics and Genomics and the Association for Molecular Pathology framework, which is the standard clinical labs use to classify genetic variants. A finding was only established as a diagnosis after qualified experts reviewed the evidence, the variant was classified as pathogenic or likely pathogenic, and a Clinical Laboratory Improvement Amendments-certified laboratory confirmed the result.
The workflow was applied to four distinct groups of previously unsolved cases:
- Neurodevelopmental conditions: 10% yield
- Rare neuromuscular disease: 6.6% yield
- Early psychosis: 13.3% yield (noting a small cohort size)
- Sudden unexpected death in pediatrics: 1% yield
In one case involving early psychosis, the model hypothesized a 22q11.2 deletion associated with DiGeorge syndrome by connecting low-quality calls on chromosome 22 with the patient’s cardiac, immune, neurodevelopmental, and psychiatric features. This variant was later confirmed through follow-up genome sequencing.
Scalability of Reanalysis
The study authors emphasize that the model did not diagnose any patient or make clinical decisions. Instead, it produced evidence-linked hypotheses for specialists to review and investigate through additional testing and clinical laboratory confirmation.
“Researchers like Catherine and me can’t possibly keep 8,000 different diseases in our heads. That’s the power of AI,” says Alan Beggs, PhD, director of the Manton Center for Orphan Disease Research, in a release.
The researchers noted that while the yield is modest, it is significant because these cases had already evaded years of expert analysis. The study suggests that expert-led periodic reanalysis could become more scalable as scientific knowledge and AI tools evolve, helping clinical teams manage the growing backlog of inconclusive genomic cases.
ID 322452409 | Ai Medical © Paradee Paradee | Dreamstime.com
