In a collaboration with a leading global pharmaceutical company, Volv Global applied machine learning to 24 million UK primary care records, surfacing approximately 1,200 likely-undiagnosed GEP-NET patients – and finding they are 5–7 years younger than those currently diagnosed.
Épalinges, Switzerland
In brief
- Approximately 1,200 UK patients have records consistent with undiagnosed GEP-NETs – roughly one additional patient for every three currently diagnosed.
- Patients flagged by the model are 5–7 years younger on average than those with confirmed diagnoses, suggesting earlier-stage detection may be achievable.
- The model achieved a ROC-AUC of 0.756 evaluated when discriminating GEP-NETs from clinically similar mimic conditions – a harder and more clinically meaningful benchmark than comparison to general population.
- The methodology is reproducible and transferable across geographies and data environments, with outputs designed to support clinician review and prospective validation.
GEP-NETs are rare malignancies whose non-specific symptoms – commonly attributed to irritable bowel syndrome, inflammatory bowel disease, or diabetes – mean patients typically wait nearly five years before receiving a confirmed diagnosis. Prognosis is closely tied to grade and stage at diagnosis; five-year survival rates for high-grade disease (G3) may be as low as 25%.
A leading global pharmaceutical company approached Volv Global to determine whether an detectable undiagnosed GEP-NET population existed within UK routine primary care records. The challenge was compounded by coding imprecision: a significant proportion of NET patients carry non-specific diagnostic codes, meaning a straightforward code-based query would systematically under-count the true population.
Volv Global applied its proprietary machine learning methodology – operating through the inTrigue framework – to the Optimum Patient Care Research Database (OPCRD), covering approximately 24 million de-identified records from around 1,100 UK GP practices. A positive cohort of 1,857 GEP-NET patients was constructed using a procedure that recovers patients not captured by direct code queries. The negative cohort was drawn from clinically relevant comparator conditions, ensuring model performance was evaluated against a meaningful real-world discrimination task.
Comprehensive phenotypic characterisation of the diagnosed cohort confirmed the multi-system burden documented in the clinical literature. Gastrointestinal, respiratory, and neurological symptoms were all significantly more prevalent in the GEP-NET cohort than in a matched random population, and treatment patterns confirmed the underrepresentation of specialist therapies in primary care records.
The patient-finding model achieved a test set ROC-AUC of 0.756 and PR-AUC of 0.427. Applied to a subset of 6.8 million patients and extrapolated across the full database, it estimated approximately 1,200 likely-undiagnosed patients at a precision of 0.85. The most predictive features were clinically coherent with the known pre-diagnostic presentation of GEP-NETs, providing a transparent basis for clinician review.
Demographic analysis found that patients flagged by the model were 5–7 years younger on average than those already diagnosed. Volv Global interprets this carefully: the age difference is a demographic observation and the hypothesis that these patients may be at an earlier disease stage would require prospective validation to confirm. No claim is made regarding clinical outcomes or treatment benefit on the basis of this finding. However, earlier detection would mean a potential significant uplift in 5-year survival rates as seen with similar cases of early detection with other cancers.
“GEP-NETs sit at the intersection of clinical complexity and data fragmentation. The non-specific symptom profile, combined with coding imprecision in primary care records, makes them a genuinely hard problem – and that is precisely where machine learning applied to population-scale real-world data can change what is possible. This work demonstrates that a meaningful undiagnosed population exists and can be characterised. The next step is prospective deployment.” – Christopher Rudolf, CEO and Founder, Volv Global
About Volv Global
Volv Global is a healthcare AI company founded in 2017 and headquartered in Épalinges, Switzerland. Its mission is to generate new knowledge at speed, close the diagnostic gap, as well as other gaps in the care pathway, to improve patient outcomes. Volv Global works across conditions where patients are difficult to detect, where the window for effective treatment is narrow, and where understanding which patients will progress or need therapy beyond standard care can meaningfully change outcomes. It is a trusted partner to leading pharmaceutical organisations across the USA and Europe, with solutions deployed in live clinical programmes.
Applying a proprietary machine learning methodology to population-scale real-world data – accessed through trusted data partners covering more than 400 million patients – Volv Global generates disease intelligence that enables pharmaceutical teams to de-risk clinical programmes, detect and stratify patient populations with greater precision, and build stronger real-world evidence. For clinicians, Volv Global’s insights are designed to surface actionable signals within existing care pathways. For patients, they translate into earlier diagnosis, better-informed treatment decisions, and a faster path to the treatment that can help them, through a diagnostic system that too often leaves difficult-to-diagnose diseases unrecognised for years.
Volv Global’s solutions each address a distinct clinical question across the patient journey, and are configured to the client’s specific research question, disease area, and healthcare setting. Volv Global does not hold patient data; all work is conducted within the governed environments operating under applicable privacy and regulatory frameworks.
Media contact
Le Vin Chin
lchin@volv.global
Volv Global SA
Route de la Corniche 3B, 1066 Épalinges, Switzerland
