AI flags about 1,200 likely undiagnosed GEP-NET patients in UK primary care

Volv Global used machine learning on 24 million UK primary care records and identified about 1,200 patients who may have undiagnosed GEP-NETs. The finding matters because the flagged patients were younger than diagnosed patients and the approach could help clinicians find rare cancers earlier.

Why it matters: - GEP-NETs are rare cancers that often present with non-specific symptoms. - Patients with these tumors commonly wait nearly five years for a confirmed diagnosis. - Earlier detection could matter because prognosis depends heavily on grade and stage at diagnosis. - High-grade disease can have five-year survival rates as low as 25%. - The work suggests machine learning could surface patients who are currently missed in routine care.

What happened: - Volv Global applied machine learning to 24 million de-identified UK primary care records. - The dataset came from the Optimum Patient Care Research Database and covered about 1,100 UK GP practices. - The project was done with a leading global pharmaceutical company. - The analysis identified approximately 1,200 likely-undiagnosed GEP-NET patients. - The flagged patients were 5–7 years younger on average than patients already diagnosed.

The details: - The model was built using Volv Global’s inTrigue framework. - A positive cohort of 1,857 GEP-NET patients was created using a procedure designed to recover patients missed by direct code queries. - The negative cohort used clinically relevant comparator conditions rather than the general population. - That setup made the task a harder and more clinically meaningful test of discrimination. - The model reached a test set ROC-AUC of 0.756 and a PR-AUC of 0.427. - Applied to a subset of 6.8 million patients and extrapolated across the database, the model estimated about 1,200 likely-undiagnosed patients. - The estimated precision was 0.85. - The strongest predictive features matched the known pre-diagnostic presentation of GEP-NETs. - Diagnosed patients showed significantly higher gastrointestinal, respiratory, and neurological symptoms than a matched random population. - Treatment patterns showed specialist therapies were underrepresented in primary care records. - Coding imprecision made simple code-based searches insufficient because many NET patients carry non-specific diagnostic codes.

Between the lines: - The age gap is a demographic signal, not proof that the flagged patients are at an earlier disease stage. - Volv Global says that hypothesis needs prospective validation. - The project is notable because it tested the model against clinically similar mimic conditions, not just healthy controls. - That design is closer to how a real clinician would face the problem in practice. - The company says the methodology is reproducible and transferable across geographies and data environments. - The output is intended to support clinician review and future validation, not to replace diagnosis.

What's next: - Volv Global says the next step is prospective deployment. - The company also expects clinician review and prospective validation of the flagged patients. - If validated, the approach could help identify more rare cancer patients earlier in routine primary care data.

The bottom line: - The analysis found a meaningful pool of likely missed GEP-NET patients in UK primary care, and the model’s results point to a practical path for earlier detection of a hard-to-diagnose cancer.