An AI Model Found Traces of Rare Genetic Diseases in Medical Records Years, and Sometimes Decades, Before Diagnosis
An AI Model Found Traces of Rare Genetic Diseases in Medical Records Years, and Sometimes Decades, Before Diagnosis
On September 11, researchers from MyOme and Mayo Clinic published a preprint describing a model that reads medical records as a chronological sequence of clinical findings. Applied to the records of nearly 3 million Mayo Clinic patients, it identified signs of ten rare genetic diseases. In a separate validation on 143 patients with prolonged undiagnosed histories, it detected the disease in every case before the diagnosis was made, with median lead times ranging from 1.9 to 21.7 years.
A patient with Fabry disease sees different doctors for years, treating pain in the hands and feet, abdominal pain, and skin rash, while no one connects these complaints into a single diagnosis. From first symptoms to a correct diagnosis of a rare disease, an average of 4 to 8 years passes, and roughly 300 million people worldwide live with such conditions. In Fabry disease the delay carries a high cost: enzyme replacement therapy can halt kidney and heart damage, but treatment cannot begin until a diagnosis is made.
The findings sit in the medical record all along. Conventional risk models compress a patient's history into a single flat list: whether something was ever noted, without regard to when. The authors divided each patient's record into quarterly intervals spanning up to 30 years, so the GRU-Attn model preserves the timing and sequence of findings. Tall stature at age 10, lens dislocation at 16, and aortic dilation at 24 are individually unremarkable. Together they describe Marfan syndrome.
One to two years before diagnosis, the temporal model distinguished future patients from those who never received the diagnosis considerably more accurately than the static model (0.928 versus 0.871 on a scale from 0.5 to 1.0). Six to ten years before diagnosis, the gap widened further: for some diseases the static model dropped nearly to chance level, while the temporal model remained substantially above it. The farther from diagnosis, the more scattered and individually weak the findings become. A pattern among them is visible only to a model that accounts for their order and pace of appearance.
In a cohort of 143 patients with the longest undiagnosed histories, the model flagged the future disease in 138 before the disease itself first appeared anywhere in their records, and in all 143 before the official diagnosis was recorded. For Marfan syndrome the median lead time was 21.7 years: the signal accumulates in the record during childhood, long before a cardiologist first examines the patient's aorta.
The temporal model's advantage is uneven. For vascular Ehlers-Danlos syndrome and Parkes Weber syndrome, where early signs are already specific, the static model performed just as well. The largest gain appeared where individual findings carry little meaning on their own and only coalesce into a diagnosis over time, as in Loeys-Dietz syndrome.
Given how rare these diseases are, even an accurate model applied to the general population generates too many false alarms per true detection to be practical. The authors propose using it as a filter within specialized clinics, where the base rate is higher: in that setting, precision for Noonan syndrome exceeded 30%, which is above the accepted threshold for low-dose CT lung screening in oncology (approximately 4%). This calculation is retrospective, drawn from a single institution, and has not been tested prospectively.
The data came from Mayo Clinic through the Mayo Clinic Platform_Accelerate program, which gives external companies access to de-identified medical records. The work is led by Matthew Rabinowitz, co-founder of Natera and chairman of MyOme. In 2003 his nephew was born with undiagnosed Down syndrome and died six days later; Rabinowitz subsequently lost a child to a genetic disease. He told Inside Precision Medicine:
"It was like being struck twice: unrelated events, but the same tragedy"
The next step is prospective validation: risk stratification from the medical record, followed by referral to a geneticist, along the lines of screening pathways already in place for hereditary cancers.