Preprint: GenEHR estimates five-year first cancer diagnosis risk from electronic health records

Preprint: GenEHR estimates five-year first cancer diagnosis risk from electronic health records

The authors evaluated GenEHR on five existing electronic health record databases. They trained a separate version of the model for each database: it estimates the probability of a first cancer diagnosis within 6–60 months and ranks individuals for targeted screening.

On September 11 the authors posted a preprint describing GenEHR. The model accounts for the sequence of diagnoses, medications, procedures, and laboratory tests, as well as the intervals between visits. A repeat test one week after an abnormal result may signal an active diagnostic workup; the same test a year later is often routine annual monitoring.

GenEHR first learns to predict the next event in a patient's history and the time until the next visit. The authors then fine-tune it for different cancer types; shared patterns across patient histories help where cases of a specific cancer are scarce. The model ranks a large population to identify who should be offered in-depth screening first.

The authors evaluated first-diagnosis risk for 8–17 cancer types, depending on the database, in a separate held-out patient cohort. In four databases, diagnoses were confirmed by cancer registries; in the fifth, the criterion was two occurrences of the same cancer code. To test for earlier signals, the authors used records ending at least three months before diagnosis.

Separately, the authors examined the one thousand individuals with the highest predicted risk. For pancreatic cancer, this group contained 34–48 times more future cases than would be expected after adjusting for age and sex, depending on the database. By their calculation, given an effective screening test, 13–18 individuals would need to be enrolled in such a program to detect one case.

In a 2025 study, part of this team built a sequential model of pancreatic cancer risk using data from the U.S. Veterans Affairs health system. The new preprint extends this approach to multiple cancer types and five databases. The authors propose first selecting individuals at elevated risk, then conducting a dedicated screening program, and initiating early treatment after cancer detection.