University of Oklahoma scientists are stepping into a high-stakes national experiment to see whether artificial intelligence can help crack some of the toughest genetic mysteries facing kids, including children with early-onset psychosis. The effort follows a closely watched research project showing that machine reasoning can surface fresh, testable leads in pediatric genetic cases that had stumped experts for years. OU leaders say the approach could cut months or even years off some families' diagnostic odysseys by helping clinicians flag promising variants and order follow-up tests sooner.
According to The Journal Record, Carlos Bustamante, the OU College of Medicine's vice dean of research, said OU is "part of a big NIH study looking at children with early psychosis." Monique Morrison, director of OU Health's Molecular Genetics Laboratory, told the paper that AI could shoulder repetitive lab tasks so that diagnostic calls stay firmly in human hands. The article also reports that OU hospitals already lean on AI tools in stroke care and radiation therapy workflows.
What the recent study showed
In June, researchers working with Boston Children's Hospital and Harvard used OpenAI's o3 Deep Research reasoning model to reanalyze 376 previously unsolved pediatric cases. After clinician review and confirmatory testing, physicians verified 18 new diagnoses - roughly a 4.8% bump in diagnostic yield, according to OpenAI and the Manton Center at Boston Children's Hospital. Instead of handing down final clinical decisions, the model generated evidence-linked hypotheses for experts to pick apart. Hospital teams stress that the tool is a research assistant meant to speed literature review and variant prioritization, not a robot doctor.
Study methods and subgroup gains
The published paper describes a clinician-led workflow that piped phenotype terms, clinical notes and filtered variant lists into the reasoning model, then asked it to rank likely explanations. That reanalysis produced new confirmed local diagnoses in 10 of 100 neurodevelopmental cases, 4 of 61 neuromuscular cases, 2 of 200 sudden unexpected pediatric death cases and 2 of 15 early-psychosis cases, for a total of 18 of 376. The full study is available in NEJM AI, and broader reviews show that periodically re-running genomic data often boosts diagnostic yields as gene-disease knowledge grows over time (PubMed Central).
How OU plans to plug in
OU has been quietly building up its genomic infrastructure and recruiting research leadership to expand clinical studies. The university's news office previously announced Bustamante as the College of Medicine's inaugural vice dean of research. The pathology directory lists Morrison among OU Health lab leaders and shows that the system runs molecular and clinical pathology services that could support periodic reanalysis pipelines. OU officials tell The Journal Record that any AI-suggested lead would still go through confirmatory testing in CLIA-certified labs before families ever see results.
Promise, limitations and next steps
Authors of the NEJM AI paper and their hospital partners are bullish but cautious. They call for prospective multicenter trials with predefined endpoints and clear calibration reporting before anyone bakes AI-assisted reanalysis into routine care. The big selling point is scale: the workflow could help teams systematically re-review old cases as new gene-disease links appear. Even so, model outputs still require human judgment and careful lab confirmation. For Oklahoma families who have waited years for answers, the OU effort is meant to speed that re-review and make access more equitable - but only under clinician oversight, rigorous validation and tight privacy safeguards.
