AI Model Helps Boston Children's Hospital Solve 18 Rare Genetic Disease Cases

For families of children with rare genetic disorders, the search for a diagnosis can stretch across years or even decades, leaving them without answers or targeted treatments. But a new collaboration between Boston Children's Hospital, Harvard University, and OpenAI is proving that artificial intelligence can meaningfully chip away at this backlog. Using OpenAI's o3 Deep Research reasoning model, researchers successfully cracked 18 pediatric cases that had previously baffled medical experts.[1][4]

The findings, published on June 18, 2026, in the peer-reviewed journal NEJM AI, detail an AI-assisted workflow designed to reanalyze the most difficult medical mysteries. The research team fed de-identified clinical and genomic data from 376 unsolved cases into the model. Following expert review and laboratory confirmation of the AI's leads, physicians established definitive diagnoses for 18 patients, representing an additional diagnostic yield of 4.8%.[1][5]

While a near 5% success rate might sound modest in other fields, in the realm of rare diseases, it represents a monumental victory. These were not fresh cases awaiting a first look; they had already been exhaustively analyzed by multidisciplinary teams and run through existing commercial diagnostic pipelines to no avail. As Dr. Catherine Brownstein, scientific director at Boston Children's Manton Center for Orphan Disease Research, noted, each of these discoveries means a definitive answer for a family that had nearly given up hope.[4][6]

The challenge of diagnosing rare diseases is as much an operational problem as it is a biological one. A patient's genome remains static, but the scientific literature surrounding it evolves daily. Researchers constantly link new genes to diseases, and laboratories frequently reclassify old variants. Human geneticists simply cannot scale the effort required to continuously cross-reference thousands of old patient files against millions of new scientific papers and fragmented databases.[2][4]

To bridge this gap, the research team did not just feed raw data into a chatbot. Instead, they assembled standardized packets for each case, utilizing Human Phenotype Ontology (HPO) terms—a controlled vocabulary for describing symptoms—alongside clinician notes and filtered variant tables. The o3 Deep Research model then synthesized this information, generating evidence-linked candidate explanations and showing its mathematical and biological work for human review.[2][4]

The AI's leads spanned several complex medical categories. The reanalysis yielded 10 new diagnoses among children with neurodevelopmental conditions, four within a neuromuscular disease cohort, two in cases of sudden unexpected pediatric death, and two among patients with early childhood psychosis.[1][3]

In one particularly striking example from the early psychosis cohort, the AI model hypothesized that a patient had a chromosome 22 deletion associated with DiGeorge syndrome. Remarkably, this specific variant was not explicitly listed in the input data provided to the model. The AI synthesized the available clinical features and existing literature to make the leap, and follow-up genome sequencing subsequently confirmed the deletion.[1]

Despite the model's advanced reasoning capabilities, strict clinical guardrails remained in place. The AI functioned strictly as a hypothesis generator, not a diagnostician. Human experts rigorously reviewed every lead using the standard clinical classification system for genetic variants, known as the ACMG/AMP framework. A finding was only counted as a diagnosis after a CLIA-certified laboratory confirmed the genetic variant and the clinical team officially returned the result to the family.[2][3]

This specific study is part of a broader, aggressive integration of artificial intelligence at Boston Children's Hospital. Beyond the 18 cases highlighted in NEJM AI, the hospital's wider use of AI tools has reportedly led to more than 40 previously unsolved rare disease diagnoses. Simultaneously, the technology has streamlined administrative burdens, saving an estimated 60,000 hours in work time and redeploying over $7 million in labor costs toward patient care.[6][8]

The success of the retrospective study paves the way for real-time clinical applications. Supported by a new grant from the OpenAI Foundation, the Manton Center is now working to develop a platform-agnostic, low-cost genetics AI copilot. The ultimate goal is to ensure that expert-led, AI-assisted reanalysis becomes a standard, scalable practice, ensuring that scientific understanding can finally keep pace with medical discovery.[4][5]