Direct comparison of diagnostic models for pancreatic cancer in patients with new-onset diabetes in the British population
8 affiliations
- 1Department of Medical and Surgical Specialties, University of Fribourg Faculty of Science and Medicine, Fribourg, Switzerland
- 2Family Medicine, Arztpraxis am Bager, Schmitten, Switzerland
- 3Basel Pharmacoepidemiology Unit, University of Basel, Basel, Switzerland
- 4Hospital Pharmacy, University Hospital Basel, Basel, Switzerland
- 5Boston Collaborative Drug Surveillance Program, Boston University, Boston, Massachusetts, USA
- 6Surgery, HFR Fribourg Cantonal Hospital, Fribourg, Switzerland
- 7Department of Visceral Surgery and Medicine, Inselspital University Hospital Bern, Bern, Switzerland
- 8University of Basel Faculty of Medicine, Basel, Switzerland
Abstract
Objective Pancreatic cancer (PC) is often preceded by elevated blood glucose levels and new-onset diabetes (NOD) may provide an opportunity for earlier detection. However, PC accounts for less than 1% of diabetes cases, making routine imaging of all patients with NOD impractical. We aimed to compare the diagnostic performance of the Enriching New-Onset Diabetes for Pancreatic Cancer (EndPAC), Boursi and Baecker models in British primary care patients with NOD and to adapt the models for use at the time of diabetes diagnosis.Design Cross-sectional analysis of a retrospective cohort.Setting We analysed routinely collected electronic primary care health records from the UK Clinical Practice Research Datalink.Participants Among 197 092 patients with NOD, 901 (0.46%) were diagnosed with PC within 3 years of the diagnosis. Included were patients aged 50–90 years with NOD diagnosed between 2002 and 2018. Inclusion also required at least one valid glycaemic measurement at diabetes diagnosis and either 3 years of follow-up or death within that period.Primary and secondary outcome measures The primary outcome was discrimination of the original and adapted models, measured by area under the receiver operating curve (AUC). Secondary outcomes included model calibration and threshold-specific diagnostic performance for the adapted Boursi model.Results In its original form, the Boursi model showed the highest performance (AUC 0.78; 95% CI 0.76 to 0.80). After adaptation, the EndPAC score achieved an AUC of 0.69 (95% CI 0.67 to 0.72), Boursi 0.76 (95% CI 0.74 to 0.78) and Baecker 0.77 (95% CI 0.75 to 0.79) in the split-sample validation cohort.Conclusions After adaptation, when restricted to parameters available at the time of NOD diagnosis, with imputed missing values, the Boursi and Baecker models performed similarly, while the EndPAC score, which uses only three variables, discriminated less well. Each model has distinct strengths; combining them may improve patient selection for further diagnostic evaluation. The adapted models were validated internally only and require external validation before clinical use. These findings support continued investigation of risk-based strategies for PC detection in primary care as part of a stepwise diagnostic pathway.
Related
- Diagnostic Value of the PEI-Q and Pancreatic Ultrasound for Identifying Reduced Pancreatic Exocrine Function in Early Type 2 Diabetes
- Biliary and Pancreatic Safety of Incretin-Based Therapies in Obesity and Type 2 Diabetes: A Propensity-Matched Real-World Comparison of Tirzepatide, Semaglutide, and DPP-4 Inhibitors
- Diabetes Among Finnish Pancreatic Cancer Patients: Prevalence, Time Interval Between Diabetes and Cancer Diagnosis, and Overall Survival.
- Evaluation of Bruch's membrane opening minimum rim width and retinal nerve fiber layer in patients with glaucoma and comparison with normal population.
- Diagnostic image quality of shoulder magnetic resonance imaging at 0.55T in healthy subjects and patients with rotator cuff injury: Prospective comparison with 1.5T.
BibTeX
@article{mellenthin2026direct,
title = {{Direct comparison of diagnostic models for pancreatic cancer in patients with new-onset diabetes in the British population}},
author = {Claudia Mellenthin and Christoph R Meier and Susan Jick and Leo Hans Bühler and Michel Adamina and Cornelia Schneider},
journal = {BMJ Open},
year = {2026},
month = nov,
doi = {10.1136/bmjopen-2026-120648},
url = {https://bmjopen.bmj.com/lookup/doi/10.1136/bmjopen-2026-120648},
note = {CC BY},
}
RIS
TY - JOUR
TI - Direct comparison of diagnostic models for pancreatic cancer in patients with new-onset diabetes in the British population
AU - Claudia Mellenthin
AU - Christoph R Meier
AU - Susan Jick
AU - Leo Hans Bühler
AU - Michel Adamina
AU - Cornelia Schneider
PY - 2026
DA - 2026/11/05
JO - BMJ Open
DO - 10.1136/bmjopen-2026-120648
UR - https://bmjopen.bmj.com/lookup/doi/10.1136/bmjopen-2026-120648
AB - Objective Pancreatic cancer (PC) is often preceded by elevated blood glucose levels and new-onset diabetes (NOD) may provide an opportunity for earlier detection. However, PC accounts for less than 1% of diabetes cases, making routine imaging of all patients with NOD impractical. We aimed to compare the diagnostic performance of the Enriching New-Onset Diabetes for Pancreatic Cancer (EndPAC), Boursi and Baecker models in British primary care patients with NOD and to adapt the models for use at the time of diabetes diagnosis.Design Cross-sectional analysis of a retrospective cohort.Setting We analysed routinely collected electronic primary care health records from the UK Clinical Practice Research Datalink.Participants Among 197 092 patients with NOD, 901 (0.46%) were diagnosed with PC within 3 years of the diagnosis. Included were patients aged 50–90 years with NOD diagnosed between 2002 and 2018. Inclusion also required at least one valid glycaemic measurement at diabetes diagnosis and either 3 years of follow-up or death within that period.Primary and secondary outcome measures The primary outcome was discrimination of the original and adapted models, measured by area under the receiver operating curve (AUC). Secondary outcomes included model calibration and threshold-specific diagnostic performance for the adapted Boursi model.Results In its original form, the Boursi model showed the highest performance (AUC 0.78; 95% CI 0.76 to 0.80). After adaptation, the EndPAC score achieved an AUC of 0.69 (95% CI 0.67 to 0.72), Boursi 0.76 (95% CI 0.74 to 0.78) and Baecker 0.77 (95% CI 0.75 to 0.79) in the split-sample validation cohort.Conclusions After adaptation, when restricted to parameters available at the time of NOD diagnosis, with imputed missing values, the Boursi and Baecker models performed similarly, while the EndPAC score, which uses only three variables, discriminated less well. Each model has distinct strengths; combining them may improve patient selection for further diagnostic evaluation. The adapted models were validated internally only and require external validation before clinical use. These findings support continued investigation of risk-based strategies for PC detection in primary care as part of a stepwise diagnostic pathway.
SN - 2044-6055
SN - 2044-6055
ER -