Researchers from Boston Children’s Hospital, Harvard University and OpenAI confirmed 18 diagnoses after using an AI-assisted workflow to reanalyze previously unsolved rare disease cases.
The study, published June 18 in NEJM AI, involved 376 cases spanning neurodevelopmental disorders, neuromuscular diseases, early psychosis and sudden unexpected pediatric deaths that had remained unresolved despite prior genetic testing, specialist evaluation and review through existing diagnostic pipelines.
Researchers used OpenAI’s o3 Deep Research model to analyze deidentified clinical and genomic data and generate evidence-based hypotheses for expert review. After additional evaluation, testing and clinical confirmation, physicians established diagnoses in 18 cases, representing an additional diagnostic yield of 4.8%.
Researchers said the findings suggest AI-assisted reanalysis may help experts generate leads when revisiting difficult cases as scientific knowledge and evidence surrounding rare diseases continue to evolve.
The highest diagnostic yield was observed in the study’s early psychosis cohort, where researchers established diagnoses in two of 15 cases. The reanalysis also led to 10 diagnoses among 100 neurodevelopmental cases, four diagnoses among 61 neuromuscular disease cases and two diagnoses among 200 sudden unexpected pediatric death cases.
Several cases illustrated the model’s ability to synthesize information from genomic data, clinical features and existing evidence during reanalysis. In one early psychosis case, the model hypothesized a chromosome 22 deletion associated with DiGeorge syndrome despite the variant not being explicitly listed in the input data. Follow-up genome sequencing later confirmed the deletion, according to the study.
Researchers also reported that the model occasionally identified more complex genetic explanations involving multiple genes. In addition, it generated a testable hypothesis involving a potential genetic mechanism for vitiligo, though researchers said further experimental validation is needed.
Of the 18 diagnoses, seven were rediscoveries of diagnoses that had already been established outside the local research workflow but were not reflected in the records reviewed by investigators. Researchers said the finding highlights challenges associated with integrating information across clinical records, databases and healthcare systems.
The AI model was not used to diagnose patients or make clinical decisions, according to the study. Instead, it generated evidence-linked hypotheses that underwent specialist review and confirmation through established clinical processes, including laboratory testing. Researchers noted that every diagnosis was made through qualified expert review and clinical confirmation.
The study also identified several limitations, including its retrospective design and the fact that researchers did not measure time savings, clinician workload, cost or effects on patient care. The authors also noted that large language models can generate plausible explanations that do not hold up under further scrutiny, making expert review essential.
The authors called for prospective, multicenter studies to compare AI-assisted genomic reanalysis with standard approaches and evaluate its effects on diagnostic yield, clinician effort, cost and patient outcomes.
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.