# paper.pdf ## Metadata - PDFFormatVersion=1.7 - IsLinearized=false - IsAcroFormPresent=false - IsXFAPresent=false - IsCollectionPresent=false - IsSignaturesPresent=false - Creator=Adobe Acrobat (64-bit) 26.1.21691 - Title=Enriching New-onset Diabetes for Pancreatic Cancer (ENDPAC): External validation using English sentinel network - Producer=3.0.38 (5.1.21) - ModDate=D:20260701073301+02'00' - CreationDate=D:20260701104051+05'30' - xmp:modifydate=2026-07-01T10:48:28+05:30 - xmp:createdate=2026-07-01T10:40:51+05:30 - xmp:metadatadate=2026-07-01T10:48:28+05:30 - xmp:creatortool=Adobe Acrobat (64-bit) 26.1.21691 - dc:format=application/pdf - xmpmm:documentid=uuid:5398c535-8691-4828-b246-d2a0c4fae02a - xmpmm:instanceid=uuid:301d1c35-d65a-43fd-8e22-a0a77e4f4e3e - pdf:producer=Adobe Acrobat (64-bit) 26.1.21691 ## Contents ### Page 1 Accepted ManuscriptBritish Journal of General PracticeEnriching New-onset Diabetes for Pancreatic Cancer (ENDPAC): External validation using English sentinel network Price, Claire; Claridge, Hugh; de Lusignan, Simon; Khalaf, Natalia; Mold, Freda; Smith, Nadia; Winn, Martyn; Lemanska, Agnieszka DOI: https://doi.org/10.3399/BJGP.2024.0253 To access the most recent version of this article, please click the DOI URL in the line above. Received 02 May 2024 Revised 06 May 2026 Accepted 27 May 2026 © 2026 The Author(s). This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License (http://creativecommons.org/licenses/by/4.0/). Published by British Journal of General Practice. For editorial process and policies, see: https://bjgp.org/authors/bjgp-editorial-process-and-policies When citing this article please include the DOI provided above.Author Accepted ManuscriptThis is an ‘author accepted manuscript’: a manuscript that has been accepted for publication in British Journal of General Practice, but which has not yet undergone subediting, typesetting, or correction. Errors discovered and corrected during this process may materially alter the content of this manuscript, and the latest published version (the Version of Record) should be used in preference to any preceding versions ### Page 2 Enriching New-onset Diabetes for Pancreatic Cancer (ENDPAC): External validation using English sentinel networkAuthorsClaire A. Price* 0000-0003-3060-9821, MSc (Open) Medicinal Chemistry) MSc (Open) Science (focussed on neurological imaging) PG Cert (University of London) Data Science BSc (Open) Natural Sciences – Biochemistry) PG Dip (Open) Mental Health Science PG Cert (Open) Social Sciences BSc (Open) Physical Sciences BA (Open) Mathematics Post-graduate researcher at University of Surrey, School of Health Sciences, Kate Granger Building, 30 Priestley Road, Surrey Research Park, Guildford, Surrey, GU2 7YH, UK Post-graduate researcher at National Physical Laboratory, Hampton Road, Teddington, Middlesex, TW11 0LW, UK Hugh Claridge 0000-0001-5998-2860, BA, MA (Cantab), Archaeology and Anthropology Post-graduate researcher at University of Surrey, School of Health Sciences, Kate Granger Building, 30 Priestley Road, Surrey Research Park, Guildford, Surrey, GU2 7YH, UK Post-graduate researcher at National Physical Laboratory, Hampton Road, Teddington, Middlesex, TW11 0LW, UK Prof Simon de Lusignan 0000-0002-8553-2641, BSc (University of London) Biochemistry MB BS, (University of London) Medicine Professor of Primary Care & Clinical Informatics at University of Oxford, Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, OX2 6GG, UK Vice Chancellor's Fellow and Professor of Primary Care & Clinical Informatics at University of Surrey, School of Health Sciences, Kate Granger Building, 30 Priestley Road, Surrey Research Park, Guildford, Surrey, GU2 7YH, UK ### Page 3 Dr Natalia Khalaf 0000-0002-7585-6183, MD (Baylor College of Medicine) Assistant Professor at Baylor College of Medicine, Center for Innovations in Quality, Effectiveness, and Safety (IQuESt), Michael E. DeBakey Veterans Affairs Medical Center, Baylor College of Medicine Department of Medicine, Houston, TX, USA Dr Freda Mold 0000-0002-6279-5537, BSc (Kingston University) Sociology PhD (University of Surrey) Sociology Senior Lecturer at University of Surrey, University of Surrey, School of Health Sciences, Kate Granger Building, 30 Priestley Road, Surrey Research Park, Guildford, Surrey, GU2 7YH, UK Dr Nadia A. S. Smith 0000-0002-6503-3001, BS (Universidad Complutense de Madrid) Mathematics MS (Universidad Complutense de Madrid) Applied Mathematics PhD (Universidad Complutense de Madrid) Applied Mathematics Principal Computer Scientist (Scientific Computing) at Royal Surrey NHS Foundation Trust, Royal Surrey NHS Foundation Trust, Scientific Computing, Egerton Road, Guildford, GU2 7XX, UK Lead Consultant (Medical Physics) at TUV SUD Nuclear, TÜV SÜD UK, Chadwick House, Warrington Road, Birchwood Park, Warrington, WA3 6AE, UK Dr Martyn Winn 0000-0003-0496-6796, BA (University of Cambridge) Natural Sciences D.Phil (University of Oxford) Theoretical Chemistry Group Leader for the Computational Biology Group, Scientific Computing Department, Science and Technology Facilities Council, Research Complex at Harwell, Didcot OX11 0FA, UK Dr Agnieszka Lemanska 0000-0003-4849-2430, BSc (Medical University of Warsaw) Pharmacy MSc (Medical University of Warsaw) Pharmacy PhD (University of Bristol) Data Science, Centre for Chemometrics Assistant Professor at Imperial Clinical Trials Unit, School of Public Health, Imperial College London, Stadium House, 68 Wood Lane, London, W12 7RH, UK *claire.a.price@surrey.ac.uk ### Page 4 UK Validation of ENDPAC Page 1 of 23AbstractBackgroundPancreatic cancer has very low survival due to late diagnosis. Symptoms are often non-specific, complicating early detection in primary care. The Enriching New-Onset Diabetes for Pancreatic Cancer (ENDPAC) algorithm uses weight change, glycaemic control, and age at diabetes onset to identify new-onset diabetes (NOD) patients at increased pancreatic cancer risk. It was developed in the USA and has not been validated in the UK.AimTo validate ENDPAC in a UK primary care population and assess its predictive utility.Design and settingRetrospective cohort study using ORCHID, a national primary care sentinel network.MethodAdults aged ≥50 with NOD and requisite glycated haemoglobin (HbA1c) and weight data were included. ENDPAC scores were calculated. Model performance was evaluated via discrimination, calibration, sensitivity, specificity, PPV and NPV. The Youden index identified optimal cutoffs. Sensitivity analyses assessed measurement timing, repeat HbA1c testing and multiple values.ResultsAmong 70,050 individuals, 185 (0.26%) developed pancreatic cancer. Cases were older with higher HbA1c and greater weight loss at diagnosis. ENDPAC achieved an area under the curve (AUC) of 0.733. An optimal cutoff of ≥3 classified 27.6% of individuals as high-risk, with 62.6% sensitivity, 72.3% specificity, 0.6% PPV and 99.9% NPV. Sensitivity analyses showed similar performance across measurement windows and handling of multiple values.ConclusionENDPAC shows moderate discrimination in UK primary care. Although it has a relatively low PPV (0.6%), integration into routine systems could provide scalable, low-cost automated risk stratification, identifying people with NOD at higher pancreatic cancer risk as part of a sequential diagnostic pathway.Key wordsPancreatic neoplasms; Population surveillance; Early diagnosis; Primary health care; Risk factor scoresHow this fits inPancreatic cancer is often a devastating disease which is hard to diagnose. An algorithm called ENDPAC has been developed in the United States to help clinicians identify people at risk of having undiagnosed pancreatic cancer. These people can be referred for an imaging investigation to diagnose or rule out cancer. We have validated ENDPAC on UK primary care data which is an important first step to enable it to be used in clinical practice in the UK.Short sentence summaryThis study validates ENDPAC in UK primary care, showing moderate discrimination and moderate PPV for pancreatic cancer risk. ### Page 5 UK Validation of ENDPAC Page 3 of 23IntroductionPancreatic cancer is often a devasting disease due to its high mortality rate with a very low 5-year survival rate of 3 to 15% [1–6]. It is the 14th most common cancer globally, accounting for 3% of all new cancer cases, with nearly half a million cases worldwide [7]. The high mortality of pancreatic cancer means that it is the fourth major cause of cancer mortality in the world [8]. The need for prompt diagnosis and early detection of pancreatic cancer is widely recognised [9–11]. Early detection is the focus of a recent NHS drive to improve the current very low survival rates [12], although even with improved detection, advances in treatments will also be required to improve survival [13]. However, there are currently no widely available screening or biomarker tests for pancreatic cancer and diagnosis depends upon presentation with symptoms to a General Practitioner (GP). This is challenging as a full-time GP will only see one case every five years [14] and especially difficult as pancreatic cancer often presents with non-specific symptoms. Symptoms also vary between patients, making diagnosis challenging for clinicians [14–16]. Therefore, it is vital to provide GPs with innovative and practical ways to facilitate early detection. This includes data-driven approaches to flag people who are at high-risk of having pancreatic cancer for further investigation. People aged over 50 with new-onset diabetes (NOD) have the highest risk of developing non-familiar or sporadic pancreatic cancer [17,18]. Diabetes is a recognised risk factor [19–21] and 30 to 65% of patients with pancreatic cancer develop diabetes [22–25]. The risk of developing pancreatic cancer is greatest within three years of NOD at ~1% [17,18,23,26,27], which represents a 6 to 8-fold increase over the general population [17,18,28,29]. These facts led to the development of a simple but promising data-driven algorithm for the earlier detection of pancreatic cancer in the United States (US). The Enriching New-onset Diabetes for Pancreatic Cancer (ENDPAC) algorithm [20] uses changes in weight and blood glucose to calculate someone’s risk of having pancreatic cancer. The aim of the algorithm is to increase or “enrich” the probability of finding individuals at high-risk of having pancreatic cancer within the larger population of people with NOD. Its clear clinical rationale means ENDPAC is easily explainable to GPs and primary care teams as well as patients. Its relative simplicity and use of measures routinely collected in primary care make ENDPAC well suited for implementation in the UK primary care setting. One significant advantage of ENDPAC is that it targets early detection as it is based on glucose and weight changes. In contrast, other data-driven risk assessment algorithms [15,19,30–36] incorporate symptoms like back pain and jaundice, which typically manifest in the later stages of pancreatic cancer. Given that pancreatic cancer is often diagnosed too late for effective treatment, leading to high mortality rates, the ability of ENDPAC to identify cases earlier in the disease progression could potentially save lives. Furthermore, ENDPAC stands out for its use of a relatively small number of routinely collected measurements, addressing concerns about data availability that arise with other algorithms. Data-driven algorithms, such as electronic Risk Assessment Tools (eRATs) [37] and QCancer [38] are showing promise in aiding clinical decisions and expediting referrals [39–41]. Before ENDPAC can be added to these tools in clinical practice, it requires validation in the UK. ENDPAC was developed in the US and validation in the UK is necessary due to significant differences between the healthcare systems and measurement standards for healthcare records of the two countries [42–45]. Furthermore, previous validations of ENDPAC [35,46,47] have also included secondary care data. Validation of the ENDPAC algorithm is necessary for its application in UK primary care due to how blood glucose is commonly recorded (ENDPAC was developed in the US using fasting blood glucose but in UK primary care, glycated haemoglobin (HbA1c) is recorded and used to diagnose diabetes). ### Page 6 UK Validation of ENDPAC Page 4 of 23 Therefore, validation specific to primary care settings in the UK is essential to ensure the algorithm's suitability for such data. The aim was to validate the ENDPAC algorithm using UK primary care data. We extracted a cohort of people with NOD and their data on NOD and pancreatic cancer diagnosis. We calculated the ENDPAC score at NOD and reported model performance, including predictive value, sensitivity and specificity.MethodsStudy design and data sourceThis a retrospective cohort study of adults with NOD between January 2006 and December 2017. They were followed up for 3 years after their first elevated HbA1c reading. We used 2006 as the beginning of the study period because coding of metabolic markers such as weight and HbA1c in primary care improved significantly from 2004 with the introduction of the Quality and Outcomes Framework and the pay-for-performance incentive scheme [48,49]. Therefore the 2006 cut-off was set to improve data completeness for at least 2 years before the index date (defined below). The dataset was censored in May 2021, so using 2017 as the end of the study period ensured there was 3 years of follow up available for each participant. Data from the Oxford-Royal College of General Practitioners (RCGP) Clinical Informatics Digital Hub (ORCHID) database was used [50]. This is a nationally representative database [51] which downloads electronic healthcare records from GP practices which belong to the RCGP Research and Surveillance Centre (RSC). The network comprises 1,800 (~27%) practices serving more than 15 million people in England and Wales. In May 2021, using the systematized nomenclature of medicine clinical terminology (SNOMED CT) system [52], we curated a cohort of people with NOD. We extracted a dataset that included demographic information, ages at diagnoses of diabetes and pancreatic cancer (where present), as well as all measurements of weight and HbA1c recorded in 2001 or later.Diabetes StatusIn the UK two elevated glycaemic values, are required to diagnose diabetes. However, in this study as in other ENDPAC studies [35,46,47], NOD was defined from the first elevated HbA1c measure (≥48mmol/mol). This was set as the index date for the study and encompasses new-onset biochemically- defined hyperglycaemia and diabetes (glycaemic onset). In the UK, glycaemic control is most commonly measured with HbA1c in mmol/mol (IFCC) [53] and this test and units were utilised in this study.ENDPACWe replicated the ENDPAC model [20] which combines age at NOD with changes in blood glucose and weight over the year before NOD to produce a risk score for having pancreatic cancer. The final ENDPAC score consists of three sub-scores: A (Δ blood glucose category) + B (Δ weight category) + C (age category at diabetes diagnosis, Table 1). The total ENDPAC score is categorised in one of three outcomes: low- (<0), intermediate- (1-2), or high-risk (≥3) of having pancreatic cancer.Table 1: ENDPAC score parameters with blood glucose (HbA1c) categories equivalent to fasting blood glucose categories presented, based on American Diabetes Association blood glucose levels for healthy, prediabetes, diabetes and uncontrolled diabetes. NOD: New-Onset Diabetes, Δ represents the change in the measurement. Table adapted from Sharma et al. [20]Fasting Blood Glucose (FBG) categories HbA1c Blood Glucose (HbA1c BG) categoriesΔ BG Category score (NOD – 1y): A ### Page 7 UK Validation of ENDPAC Page 5 of 23 BG range (mg/dl) BG range (HbA1c mmol/mol) Score Score range BG category at ~ 1 years < 100 < 42 1≥ 100 to < 110 ≥ 42 to < 44 2≥ 110 to < 126 ≥ 44 to < 48 3 BG category at glycaemically-defined NOD≥ 126 to < 160 ≥ 48 to < 60 4≥ 160 ≥ 60 5 +1 to +4Δ Weight category Δ Weight score: BΔ Weight (kg) Score Score range≤ -6.0 +6 > -6.0 to ≤ -4.0 +4 > -4.0 to ≤ -2.0 +2 > -2.0 to < +2.0 0≥ +2.0 to < +4.0 -2≥ +4.0 to < +6.0 -4≥ +6.0 -6 -6 to +6Age at glycaemically-defined NOD Age score: CAge Range Score Score range < 60 -1≥ 60 to < 70 0≥ 70 +1 -1 to +1Total Score A + B + CStudy CohortStudy participants were aged 50 years or older, with NOD diagnosed between 2006 and 2017. Eligible patients were required to have: 1. An HbA1c ≥48 mmol/mol reading at index date preceded by at least one HbA1c <48 mmol/mol reading in the past 3-24 months 2. No HbA1c ≥48 mmol/mol reading before the index date 3. No history of pancreatic cancer before the index date 4. Weight on the index date (± 30 days) 5. Weight prior to the index date (between 10 years and 3 months prior) If multiple readings for weight or HbA1c readings were available, the flowchart by Sharma et al. [20] was followed to select which readings were used to calculate the ENDPAC score.Statistical analysisDescriptive StatisticsBaseline demographic statistics (sex, ethnicity and index of multiple deprivation [IMD] quintile) were analysed with descriptive statistics, means ± standard deviations (SD), counts and percentages. All counts were rounded to the nearest five as per the rules of statistical disclosure.PredictorsThe study included three risk factors, as defined by the ENDPAC model. These were: age at index date, change in body weight (kg), and change in HbA1c (mmol/mol). ### Page 8 UK Validation of ENDPAC Page 6 of 23Outcome identificationThe outcome was the diagnosis of pancreatic cancer within three years after the index date. Pancreatic cancer diagnosis was defined based on SNOMED CT clinical coding system, using the code 363418001 for malignant tumor of pancreas (disorder) and all sub-codes, excluding codes related to cancers that have metastasised to the pancreas (94164003, 94212002, 94325008, 94354002, 94459006, 9460001, 944618007, 285614004) [52].Missing dataThis study was a validation of the ENDPAC model. ENDPAC scores were only calculated for people who had the required information (complete cases).Model validationThe positive predictive value (PPV), negative predictive value (NPV), sensitivity and specificity were calculated for ENDPAC scores from ≤0, through 1 to 5, and ≥6. The percentage to be referred for further investigation at each score was calculated as the number of people with a score greater than or equal to the score divided by the whole cohort population times 100. Calibration (predicted probability matches with observed outcomes) was assessed using calibration plots of the slope of the observed proportion of events against the predicted risk. The model’s discrimination (ability to distinguish between those with and without pancreatic cancer) was assessed by area under the receiver operating characteristic curve (AUC). The optimal ENDPAC cutoff value (set to identify people at high-risk of pancreatic cancer, previously set at ≥3 during model derivation) was determined, using the Youden index, which maximises both the sensitivity and specificity of the model [54]. The performance of the model was compared with other published models for the detection of pancreatic cancer using the Chi-square test for the models’ sensitivity and specificity. To account for multiple comparisons, the statistical significance was set at p < 0.01.Subgroup analysesRisk score stratification was utilised to explore the differences in cancer rates between the risk score categories (low-, intermediate- and high-risk). Model sensitivity was also investigated for different time windows from index date (3-6 months, 6-9 months, 9-15 months, 15-18 months, 18-21 months, 21-24 months and 24 months-10 years) of HbA1c and weight measurements utilised. Sensitivity analysis was conducted to investigate whether the predictions changed when using highest, lowest or mean readings where multiple weight or HbA1c readings were available. The effect of using a single elevated HbA1c reading as the definition of diabetes instead of requiring people to have a second HbA1c reading within 90 days was also investigated.Software and reproducibilityThe database was managed in Structured Query Language (SQL) Server Management Studio version v18.9.1. Data analyses were conducted using RStudio version 2023.03.0+386 "Cherry Blossom".ResultsA total of 70,050 people from 669 GP practices met the inclusion criteria (Figure 1). We initially identified individuals with a first elevated HbA1c reading (box a), excluded those aged <50 years or ### Page 9 UK Validation of ENDPAC Page 7 of 23 with prior pancreatic cancer (box b), then excluded those without a reading for normal HbA1c in the 3 to 24 months before index (box c) and then those without two weight measurements required to calculate the ENDPAC score (box d). Participants in the final cohort had a mean age of 68.0 ± 9.9 years and 57.2% were male (Table 2). Pancreatic cancer occurred in 185 individuals (0.26%).Figure 1: Consort diagram of cohort population.Table 2 summarises the demographic and clinical characteristics of the whole cohort and for participants with and without pancreatic cancer. Participants with pancreatic cancer were older (72.4 ± 8.4 vs. 68.0 ± 9.9 years, p < 0.001) at index date than those without. They also had higher HbA1c at index date (57.0 ± 15.8 vs. 53.2 ± 9.2 mmol/mol, p < 0.001), greater HbA1c increase, lower weight at index date (82.1 ± 19.8 vs. 88.0 ± 19.9 kg, p < 0.001), and greater weight loss. ### Page 10 UK Validation of ENDPAC Page 8 of 23Table 2: Summary of the demographics and clinical characteristics for the whole cohort, and separately for participants with and without pancreatic cancer. All counts were rounded to the nearest five to comply with the rules of statistical disclosure.Whole cohort (n = 70,050) People with Pancreatic Cancer (n = 185) People without Pancreatic Cancer (n = 69,865)Age, mean (SD) 68.0 (9.9) 72.4 (8.4) 68.0 (9.9) Sex, n (%) Females Males 29,975 (42.8) 40,075 (57.2) 75 (40.5) 110 (59.5) 29,900 (41.5) 39,965 (57.2) Ethnicity, n (%) Asian Black Mixed Other White Missing 3,670 (5.2) 1,760 (2.5) 360 (0.5) 330 (0.5) 54,220 (77.4) 9,710 (13.9)5 (2.7)5 (2.7)5 (2.7) 5 (2.7) 140 (74.9) 40 (21.4) 3,670 (5.3) 1,750 (2.5) 360 (0.5) 330 (0.5) 54,090 (77.4) 9,670 (13.8) Index of Multiple Deprivation, n (%) 1 (least deprived) 2 3 4 5 (most deprived) Missing 13,035 (18.6) 13,240 (18.9) 13,975 (20.0) 14,205 (20.3) 13,355 (19.1) 2,240 (3.2)30 (16.2) 40 (21.4) 35 (18.9) 45 (24.3) 35 (18.9)10 (5.4) 13,010 (18.6) 13,200 (18.9) 13,940 (20.0) 14,165 (20.3) 13,315 (19.1) 2,235 (3.2) HbA1c at index, mean (SD), mmol/mol 53.2 (9.3) 57.0 (15.8) 53.2 (9.2)HbA1c before index, mean (SD), mmol/mol 44.2 (3.0) 43.5 (3.2) 44.2 (3.0) Body weight at index, mean (SD), kg 88.0 (19.9) 82.1 (19.8) 88.0 (19.9) Body weight before index, mean (SD), kg 86.8 (19.1) 84.6 (20.1) 86.8 (19.1) Table 3 presents ENDPAC sub-scores for age, HbA1c and weight and total scores at index date, for the whole cohort and for participants with and without pancreatic cancer. ENDPAC score was significantly higher in participants with pancreatic cancer than those without (4.10 ± 3.57 vs. 1.04 ± 3.34, p < 0.001), with a striking difference in a higher weight sub-score (1.61 ± 3.28 vs. −0.77 ± 3.10, p < 0.001).Table 3: Mean ENDPAC scores and subscores with significance (p-value) of the difference between people with and without pancreatic cancerENDPAC Scores and subscores, mean (SD) Whole cohort (n = 70,050) People with Pancreatic Cancer (n = 185) People without Pancreatic Cancer (n = 69,865) p-valueAge HbA1c (BG) Weight Total 0.22 (0.79) 1.59 (0.83) -0.76 (3.11) 1.04 (3.35) 0.59 (0.63) 1.89 (0.94) 1.61 (3.28) 4.10 (3.57) 0.21 (0.79) 1.59 (0.83) -0.77 (3.10) 1.04 (3.34) <0.001 <0.001 <0.001 <0.001 The AUC of the ENDPAC model (Figure 2a) was 0.733. Calibration plots showed a slope of 0.256, intercept of -0.003 and a Brier score of 0.265 (Figure 2b). When calibration plots were repeated just around decision thresholds the metrics were largely unchanged (slope of 0.27, intercept of -0.004 and Brier score of 0.249). ### Page 11 UK Validation of ENDPAC Page 9 of 23Figure 2: Model performance plots: a) Area Under the Receiver Operating Characteristic (ROC) Curve, b) Calibration plotTable 4 presents the numbers of participants at each ENDPAC score, sensitivity, specificity, PPV and NPV. The proportion of individuals from the whole cohort who would be flagged as requiring referral for cancer investigation if the score was used as the cutoff for high-risk was also presented as well as the Youden Index used to define the optimal cutoff. The highest number of participants, 29,240 people (41.7%) had ENDPAC score of ≤0 and using cutoffs from the original model derivation would be classified as low risk. A total of 21,315 people (30.4%) had ENDPAC scores of 1 and 2 and would be classified as intermediate risk, and 19,500 people (27.8%) had ENDPAC scores ≥3 and would be classified as high-risk. The Youden index was the highest at 0.35 for the ENDPAC score 3 identifying this as the optimal threshold and confirming it as the cutoff for high-risk, achieving sensitivity of 62.6%, specificity 72.3%, PPV 0.6%, and NPV 99.9%. Using this threshold, 27.8% of individuals would have ENDPAC score of ≥3 and would be referred for further investigation for pancreatic cancer. This is equivalent to approximately 29 referrals per practice over the 12-year study period, or about 2 to 3 referrals per practice per year. If data completeness increases (e.g. through routine recording of weight and/or HbA1c levels) these referral numbers are likely to increase. ### Page 12 UK Validation of ENDPACP a g e 10 | 23Table 4: Number of people (n and % of the whole cohort) with an ENDPAC score, sensitivity, specificity, PPV, NPV and percentage of the whole cohort who would require referral if the score was used as the cutoff for high-risk. The optimal cutoff of ≥3 (as determined by the highest Youden Index) was highlighted in bold. All counts were rounded to the nearest five to comply with the rules of statistical disclosure. *As the table is truncated with scores ≤0 grouped together a percentage of people to be referred cannot be calculated as this would include the whole cohort.Risk score Number of people, n (%) Number to be referred if the score was used as a cutoff, n (%) Number of pancreatic cancer cases, n (% of the people with the score) Sensitivity (%) Specificity (%) PPV (%) NPV (%) Youden Index≤0 29,240 (41.7) * 25 (13.5) 14.4 58.2 9.23 99.5 0.201 8,595 (12.3) 40,815 (58.3) 15 (8.1) 85.6 41.7 0.39 99.9 0.272 12,720 (18.2) 32,220 (46.0) 25 (13.5) 77.0 53.9 0.45 99.9 0.313 5,410 (7.7) 19,500 (27.8) 15 (8.1) 62.6 72.1 0.60 99.9 0.354 5,030 (7.2) 14,090 (20.1) 15 (8.1) 53.5 79.9 0.71 99.8 0.335 2,030 (2.9) 9,060 (12.9) 15 (8.1) 44.4 87.1 0.92 99.8 0.32≥6 7,030 (10.0) 7,030 (10.0) 70 (37.8) 37.4 90.0 1.00 99.8 0.27 ### Page 13 UK Validation of ENDPACP a g e 11 | 23The mean time after index date to pancreatic cancer diagnosis was 416 days (SD 308; median 379.5; IQR 145.25 - 669.25). Of these cases, 25 (13.5%) occurred in the low-risk group, 45 (24.3%) in the intermediate-risk group and 115 (62.6%) in the high-risk group. The prevalence of pancreatic cancer in the high-risk group was 0.6%, representing a 2.2-fold enrichment from the whole cohort prevalence of 0.26%.Sensitivity analysisMeasurement lead timesThe majority of people had HbA1c (41,615, 59.4%) and weight (40,845, 58.3%) readings taken in the 9 to 15 months before index date (Supplementary Table 1). Sensitivity was highest when HbA1c measurements were taken 9 to 15 months pre-index (65.1% sensitivity; 75.6% specificity). For weight readings, sensitivity was highest for readings taken 18 to 21 months pre-index (66.7%; specificity 84.2%).Second elevated HbA1c within 90 daysOf the 70,050 people, 11,110 (15.9%) had a second elevated HbA1c reading within 90 days of their first elevated result; 55 of these individuals (0.5%) were diagnosed with pancreatic cancer. When the analysis was restricted to this subgroup, the model’s AUC was 0.741 and was not significantly different from the main model (p = 0.76). The optimal cutoff identified by the Youden index (0.37) was an ENDPAC score of 4 yielding a sensitivity of 61.1%, specificity of 76.3%, PPV of 1.2%, and NPV of 99.8%. For comparison, using the whole cohort cutoff of 3 for this subset of people yielded a sensitivity of 68.5%, specificity of 68.8%, PPV of 1.1%, and NPV of 99.8%.Multiple HbA1c valuesAfter selecting HbA1c readings according to the time windows specified in Sharma et al. [20], 1,255 (1.8%) people had multiple eligible HbA1c readings. When the analysis was repeated using different approaches to these multiple values, the AUCs were nearly identical and not significantly different (p = 0.97): 0.734 when using the highest HbA1c, 0.732 when using the lowest, and 0.732 when using the average (mean). The Youden index (0.35 for all analyses) identified an optimal cutoff of 3 across all approaches. At this cutoff, the three models produced the same sensitivity (62.6%), PPV (0.6%) and NPV (99.9%) with specificity differing marginally (72.5%, 72.1% and 72.3%, respectively).Comparison with other modelsThe chi-squared test comparing AUCs showed no statistically significant difference between our results and those obtained for ENDPAC in other cohorts (Supplementary Table 2). However, when compared with other models, significant differences in AUC were observed for the studies by Boursiet al. [32] (p < 0.001), and Malhotra et al. [19] (p = 0.002). Although the QCancer (Pancreas) [14] model did not differ significantly from ENDPAC in terms of sensitivity (p = 0.52), there was a significant difference in specificity and AUC (p < 0.001). QCancer (Pancreas) had higher specificity and AUC, indicating potentially increased accuracy, but also utilised imputation for missing data so this may not reflect a clinically meaningful difference. Insufficient data were available to compare sensitivity and specificity for other models. ### Page 14 UK Validation of ENDPACP a g e 12 | 23DiscussionSummaryIn this large, nationally representative UK primary care cohort of more than 70,000 adults with NOD, we validated the ENDPAC model and found performance consistent with previous studies [35,46,47]. ENDPAC demonstrated moderate discrimination (AUC 0.733), supporting its role as a risk- stratification aid rather than a standalone screening or diagnostic tool [55]. However, PPV was relatively low and calibration was poor, with a shallow slope indicating overestimation of absolute risk. Sensitivity analyses showed robust performance across HbA1c measurement timing, handling of multiple values and paired hyperglycaemic readings. The baseline prevalence in our cohort (0.26%) was lower than in prior validation studies (0.7 to 0.8%), although a similar ~2.2-fold enrichment was observed. Consequently, the PPV was lower, reflecting either a lower underlying population incidence or how cancers are reported in primary care [56]. However, ENDPAC effectively increased or enriched risk prediction, increasing the probability of finding individuals with NOD at high-risk of pancreatic cancer. Individuals with a score≥3 accounted for 62.6% of cancer cases which more than doubled the prevalence (0.6%) compared with the overall cohort of people with NOD, equating to an expected 2-3 additional referrals per practice annually.Strengths and limitationsA key strength of this study is the use of a large, nationally representative [51] primary care dataset, enabling real-world evaluation of ENDPAC. Inclusion of practices within a vetted research network ensures high data quality [49]. Defining NOD using HbA1c minimizes misclassification and reflects real-world practice, as many people lack a confirmatory reading within 90 days [57]. Comprehensive sensitivity analyses demonstrated robust and consistent discrimination across measurement timing and multiple values. Stratified analyses further allow evaluation of performance across patient subgroups, informing potential refinement of risk thresholds. Limitations include incomplete and opportunistic recording of weight and HbA1c [58], with variable measurement intervals that may particularly affect the weight-loss component. Pancreatic cancer diagnoses were extracted from primary care records, rather than cancer registry linkage, potentially missing or miscoding cases [59–61], although primary care generally received reliable cancer notifications [62]. Calibration was limited, which is expected in low-prevalence settings, where limited event counts can constrain the reliability of calibration assessment, particularly in external validation contexts [63,64]. However, a model with good discrimination may retain clinically useful ranking ability even in the presence of imperfect calibration [65]. Additionally, some patients may present with advanced diabetes, as the disease is frequently undiagnosed; approximately 2% of UK adults have undiagnosed diabetes, and 30% of cases of diabetes remain undiagnosed [66]. This limits the window for early cancer detection. The absence of cancer staging data also prevented assessment of performance in early versus late disease. As with all retrospective analyses, selection and information bias are possible.Comparison with existing literatureOur study represents the largest validation of ENDPAC and is reflective of data available in the UK primary care system. Consistent with the original derivation study [20], an ENDPAC score of ≥3identified high-risk patients. Performance metrics aligned with prior work [35,46,47] showing modest discrimination with AUCs between 0.70 and 0.75 and modest PPV between 1.7% and 2.6%. ### Page 15 UK Validation of ENDPACP a g e 13 | 23This supports ENDPAC’s transportability across health systems and its reproducibility and risk- enrichment capability. Routine longitudinal metabolic changes may therefore function as practical, low-cost early detection signals for pancreatic cancer with high score identifying at risk individuals several months prior to clinical presentation [20]. However, heterogeneity in sensitivity, specificity and optimal cutoff across populations (e.g. the validation by Khan et al. [46] reports an optimal cutoff ≥1) highlights the importance of validation and calibration for each population [55,67]. As observed in prior studies, low pancreatic cancer prevalence limits PPV. This reflects a fundamental epidemiological constraint: even in enriched populations, most flagged individuals will not develop cancer. This highlights the limitations of population-level screening based solely on NOD status, limiting the clinical utility of the model as a stand-alone screening tool. Nonetheless, our results support the potential role of ENDPAC as a risk-enrichment tool within a multi-step diagnostic pathway, helping clinicians prioritise a subset of newly diabetic patients for further investigation, who may have pancreatogenic hyperglycaemia that is otherwise difficult to recognise in primary care [14,68]. ENDPAC has the advantage of simplicity, relying on just three routinely recorded variables. Other models [15,19,31–34,36], achieve comparable AUCs but differ in sensitivity and specificity, illustrating trade-offs between false positives and false negatives. They also often require additional data inputs which may limit feasibility of use in routine practice. ENDPAC also differs from models such as QCancer (Pancreas) [15] which use cross-sectional symptoms, demographics and comorbidities to flag undiagnosed pancreatic cancer at the population level. In contrast, ENDPAC operates within a higher-risk NOD population, improving pre-test probability but limiting applicability to non-NOD presentations. These approaches are therefore complementary: population-level models support broad triage, while ENDPAC enriches risk within a clinically defined high-risk subgroup. Other approaches, such as those by Clift et al. [69] and the UK Biobank [70], suggests that combining clinical, biochemical and genetic data may improve risk stratification. However, missing data remains a challenge across all modelling strategies, particularly in routinely collected datasets where information is generated for clinical rather than research purposes [71,72]. While multiple imputation is often recommended, complete-case analysis remains common in external validation due to concerns about transportability of imputation models [65,73]. For ENDPAC, imputation has shown minimal benefit [58] and may introduce additional uncertainty. Even with the parsimony of ENDPAC, a score could only be calculated in 16.3% of individuals presenting with NOD. As in other validation studies [35,46,47], we omitted the original requirement for a confirmatory raised HbA1c at diagnosis, which increased the proportion of individuals eligible for scoring from 2.6% to 16.3%. Although using the smaller cohort the PPV increased to 1.2% there is little clinical utility in having a model that can only be applied to a few people [74,75]. Therefore, data completeness remains a key constraint. This reflects a broader limitation of routine data-based models, where increasing predictor complexity typically reduces usable sample size. Ultimately, sufficient routinely recorded information is required to generate meaningful predictions, which must be considered when designing early detection strategies for pancreatic cancer in primary care.Implications for research and practiceFurther research should move beyond retrospective validation to determine whether ENDPAC- guided risk stratification improves meaningful clinical outcomes such as earlier pancreatic cancer ### Page 16 UK Validation of ENDPACP a g e 14 | 23diagnosis and survival. Prospective studies, such as those being undertaken by Chari et al. [76] and Oldfield et al. [77], which include linkage to imaging and cancer registry endpoints, are needed to evaluate effectiveness, safety, and cost-benefit balance in routine care. These should be accompanied by health-economic evaluation of imaging demand, missed-case trade-offs and equity impacts, alongside implementation research assessing usability, acceptability and workflow integration, given that predictive tools often fail to translate into practice [78,79]. Evidence from primary care suggests that underuse of NICE referral guidance remains a barrier to early diagnosis [69,80] and a possible gap in care quality which contributes to diagnostic delay and poorer outcomes than many European countries [81,82]. In response, a NHS initiative launched in 2025 aims to improve referral pathways for high-risk abdominal imaging [12]. This will evaluate how NICE guidelines (which like ENDPAC focus on age and weight loss for people with NOD) are incorporated into early detection pathways. A randomised controlled trial in the US is currently testing whether referral strategies based on age, glycaemic change, and weight loss improve early detection [76]. Parallel work is assessing embedding ENDPAC-like tools directly into primary care systems [83]. This UK validation therefore represents an early step in translating model development into practice. Ultimately, successful implementation requires integration into primary care workflows [55,78,84,85], not just accurate validation and should be understood within the broader context of diagnostic strategy design. We envision ENDPAC being integrated into clinical workflows in the same way eRATs are utilised [86], to support early cancer detection by flagging patients with elevated risk based on routinely collected data. These tools, including QCancer, have demonstrated value in supporting decision-making and improving alignment with NICE NG12 referral guidance for suspected cancer [15,37,86,87]. However, they are most effective when embedded in a multi-step pathway rather than used in isolation. Clinical judgement remains essential, with risk scores acting as a triage mechanism supported by symptom review and patient history to avoid unnecessary referrals, particularly in low-incidence cancers where PPV is inherently constrained. This has implications for how thresholds are defined. A single cut-off is unlikely to be optimal; instead, stratified thresholds tailored to clinical context and patient preference are more appropriate [88]. Shared decision-making is essential, as acceptable risk thresholds vary, with some evidence supporting referral at risks as low as 1% [89] and higher thresholds (e.g. 3%) prompting urgent pathways. Decision curve analysis may help define these thresholds by balancing benefits and harms and to help assess the probability of disease at which a patient would opt for treatment [90]. Given the low incidence of pancreatic cancer (17.6 per 100,000 annually [91,92]), false positives from predictive models are unavoidable, making population-wide screening inappropriate due to downstream harms including unnecessary testing, psychological burden and cost [93,94]. However, comparable trade-offs are accepted in other cancer screening programmes, where low PPVs are offset by earlier detection benefits. For example, QCancer for gastro-oesophageal cancer produces PPVs of 0.2 to 0.8% with large numbers of false positives (only 405 cases from 53,529 women flagged as high-risk) [95]. Breast, prostate and colorectal screening programmes also routinely accept false positive rates of 10 to 60% [96]. ENDPAC’s estimated false positive rate (~99%) is also similar to low-dose CT lung cancer screening (~96%) [97]. In these contexts, benefits of earlier diagnosis are considered to outweigh harms when embedded in structured pathways [97,98]. Importantly, evidence suggests that the psychological impact of false positives on patients is generally transient and may even reduce anxiety or increase future screening engagement [96,99], although clear communication and informed consent remain essential. ### Page 17 UK Validation of ENDPACP a g e 15 | 23The potential value of earlier detection is reinforced by economic considerations. Pancreatic cancer treatment is expensive and with costs continuing to rise [100,101]. Therefore, while increased testing from false positives may strain services, earlier detection could reduce long-term treatment costs and improve outcomes by increasing the proportion of patients eligible for resection and curative treatment. For instance, estimates suggest that screening could increase resectability rates from 10–15% to around 40% and may remain cost-effective even with more modest improvements above 25% [102]. Currently, NHS spending on cancer diagnosis is ~£1bn annually [103]. This may be reduced through earlier detection reducing ongoing treatment costs but although improved referral systems have been associated with survival gains [104] robust cost-effectiveness data remain limited [105]. Maximising benefit while limiting harm therefore requires embedding ENDPAC within a multi-step diagnostic pathway. Secondary triage using clinical review, biomarkers and imaging can substantially reduce false positives. For example, Sharma et al. demonstrated that additional review could halve false positives and increase the three-year cancer risk from 3.6% to 10% in people flagged as high- risk [20]. Such pathways also improve diagnostic specificity by identifying alternative explanations such as steroid use or poorly controlled diabetes. Further refinement could improve performance and clinical utility. ENDPAC was originally based on evidence that pancreatogenic diabetes is associated with more rapid glyceamic deterioration and higher HbA1c at diagnosis than Type 2 diabetes [23,106]. Therefore, sensitivity analysis could be conducted to see if people with very high HbA1c at diagnosis (e.g. ≥60) are at increased risk or warrants further stratification of the model. More broadly, incorporating additional risk modifiers could improve targeting. Evidence supports stratification by age, sex, ethnicity, deprivation (IMD) and lifestyle factors (e.g., smoking and alcohol use), reflecting known variation in incidence and diagnostic timing (including later peak diagnosis in women and incidence variation by deprivation) [92,107,108]. Further improvements could include using BMI alongside weight loss to increase data availability and shifting from categorical to continuous variables to better distinguish risk within groups. These refinements would help target investigations for those most at risk, which is essential given the rarity of pancreatic cancer and the consequent low PPV of even highly accurate models. Alongside model refinement, additional work should examine the influence of data completeness, measurement timing and frequency as well as the relationship between diabetes onset and pancreatic cancer detection. Our findings also reinforce the need for structured recording of weight and HbA1c in primary care, which would benefit not only pancreatic cancer detection but also identification of undiagnosed diabetes. This is itself a significant public health issue, with an estimated one million people in the UK living with undiagnosed diabetes, representing nearly a quarter of cases [49,66,109]. Ultimately, routine clinical data offer opportunities for earlier detection across multiple conditions, but only where information is sufficiently complete and consistently recorded. Improving pancreatic cancer outcomes will therefore require not only better risk stratification tools and diagnostic pathways, but also advances in treatment and biomarkers, particularly given tumour heterogeneity. Within this broader framework, ENDPAC provides a structured and scalable approach to risk enrichment in primary care. ### Page 18 UK Validation of ENDPACP a g e 16 | 23In conclusion, this study provides robust, real-world validation of ENDPAC within UK primary care. ENDPAC demonstrates consistent, moderate discrimination and can effectively enrich pancreatic cancer risk among individuals with NOD. While limitations related to calibration, data completeness, and low disease prevalence restrict its use as a standalone tool, it reliably enriches a population of people with NOD to identify a high-risk subgroup for pancreatic cancer and performs consistently with existing literature. Its simplicity, reproducibility and compatibility with routine clinical data means it has clear potential as part of a structured, multi-step early detection strategy. However, prospective evaluation, integration into clinical workflows and careful management of trade-offs between early detection, false positives and resource use are needed to determine whether ENDPAC-guided strategies can meaningfully improve diagnostic timeliness and clinical outcomes. ### Page 19 UK Validation of ENDPACP a g e 17 | 23Additional InformationRegistered Report ProtocolThis study was conducted using the registered report process. The protocol can be found on Open Science Framework https://osf.io/ugqm7.Ethical ImplicationsThe study was approved by the University of Surrey Ethics Committee on 30th September 2022 (reference number: FHMS 21-22 269 EGA). Access to Research and Surveillance Centre (RSC) data has been approved by Royal College of General Practitioners (RCGP) under data request RSC_0420.Data AvailabilityData will remain under the control of the Oxford-Royal College of General Practitioners Clinical Informatics Digital Hub (ORCHID, orchid.phc.ox.ac.uk) and can be accessed following all necessary approvals.AcknowledgementsWe thank patients and practices who are members of the Oxford-RCGP RSC network. This project was funded as part of an Engineering and Physical Sciences Research Council iCase studentship undertaken by CP. The work of NPL co-authors was funded by the UK Government’s Department for Science, Innovation & Technology through the UK’s National Measurement System programmes.Competing InterestsThe authors declare that no competing interests exist. ### Page 20 UK Validation of ENDPAC Page 18 of 23References[1] Rachet B, Maringe C, Nur U, et al. Population-based cancer survival trends in England and Wales up to 2007: an assessment of the NHS cancer plan for England. Lancet Oncol 2009;10:351–69. [2] Rahib L, Smith BD, Aizenberg R, et al. Projecting cancer incidence and deaths to 2030: the unexpected burden of thyroid, liver, and pancreas cancers in the United States. Cancer Res 2014;74:2913–21. [3] Woodmansey C, McGovern AP, McCullough KA, et al. Incidence, Demographics, and Clinical Characteristics of Diabetes of the Exocrine Pancreas (Type 3c): A Retrospective Cohort Study. Diabetes Care 2017;40:1486–93. [4] Bray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2018;68:394–424. [5] Arnold M, Rutherford MJ, Bardot A, et al. Progress in cancer survival, mortality, and incidence in seven high-income countries 1995-2014 (ICBP SURVMARK-2): a population-based study. Lancet Oncol 2019;20:1493–505. [6] Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin 2023;73:17–48. [7] Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209–49. [8] Wang W, Chen S, Brune KA, et al. PancPRO: risk assessment for individuals with a family history of pancreatic cancer. J Clin Oncol 2007;25:1417–22. [9] Kenner BJ, Chari ST, Maitra A, et al. Early Detection of Pancreatic Cancer-a Defined Future Using Lessons From Other Cancers: A White Paper. Pancreas 2016;45:1073–9. [10] Smith LM, Mahoney DW, Bamlet WR, et al. Early detection of pancreatic cancer: Study design and analytical considerations in biomarker discovery and early phase validation studies. Pancreatology 2024;24:1265–79. [11] Pedro BA, Wood LD. Early neoplastic lesions of the pancreas: initiation, progression, and opportunities for precancer interception. J Clin Invest 2025;135. [12] England NHS. NHS England » NHS launches drive to catch one of the most lethal cancers 2025. https://www.england.nhs.uk/2025/06/nhs-launches-drive-to-catch-one-of-the-most-lethal- cancers/ (accessed 20 November 2025). [13] Silva HM, Rosa RCA. Pancreatic cancer in the era of precision medicine: challenges, advances, and the future of therapeutic strategies. J Egypt Natl Canc Inst 2025;37:64. [14] Evans J, Chapple A, Salisbury H, et al. “It can’t be very important because it comes and goes”— patients’ accounts of intermittent symptoms preceding a pancreatic cancer diagnosis: a qualitative study. BMJ Open 2014;4:e004215. [15] Hippisley-Cox J, Coupland C. Identifying patients with suspected pancreatic cancer in primary care: derivation and validation of an algorithm. Br J Gen Pract 2012;62:e38-45. [16] Olson SH, Xu Y, Herzog K, et al. Weight Loss, Diabetes, Fatigue, and Depression Preceding Pancreatic Cancer. Pancreas 2016;45:986–91. [17] Chari ST, Leibson CL, Rabe KG, et al. Probability of pancreatic cancer following diabetes: a population-based study. Gastroenterology 2005;129:504–11. [18] Pannala R, Basu A, Petersen GM, et al. New-onset diabetes: a potential clue to the early diagnosis of pancreatic cancer. Lancet Oncol 2009;10:88–95. [19] Malhotra A, Rachet B, Bonaventure A, et al. Can we screen for pancreatic cancer? Identifying a sub-population of patients at high risk of subsequent diagnosis using machine learning techniques applied to primary care data. PLoS One 2021;16:e0251876. [20] Sharma A, Kandlakunta H, Nagpal SJS, et al. Model to Determine Risk of Pancreatic Cancer in Patients With New-Onset Diabetes. Gastroenterology 2018;155:730-739.e3. ### Page 21 UK Validation of ENDPAC Page 19 of 23 [21] Magruder JT, Elahi D, Andersen DK. Diabetes and pancreatic cancer: chicken or egg? Pancreas 2011;40:339–51. [22] Lemanska A, Price CA, Jeffreys N, et al. BMI and HbA1c are metabolic markers for pancreatic cancer: Matched case-control study using a UK primary care database. PLoS One 2022;17:e0275369. [23] Pannala R, Leirness JB, Bamlet WR, et al. Prevalence and clinical profile of pancreatic cancer- associated diabetes mellitus. Gastroenterology 2008;134:981–7. [24] Chari ST, Leibson CL, Rabe KG, et al. Pancreatic cancer-associated diabetes mellitus: prevalence and temporal association with diagnosis of cancer. Gastroenterology 2008;134:95–101. [25] Liao W, Clift AK, Patone M, et al. Identifying symptoms associated with diagnosis of pancreatic exocrine and neuroendocrine neoplasms: a nested case-control study of the UK primary care population. Br J Gen Pract 2021;71:e836–45. [26] Andersen DK, Korc M, Petersen GM, et al. Diabetes, Pancreatogenic Diabetes, and Pancreatic Cancer. Diabetes 2017;66:1103–10. [27] Aggarwal G, Rabe KG, Petersen GM, et al. New-onset diabetes in pancreatic cancer: a study in the primary care setting. Pancreatology 2012;12:156–61. [28] Huang BZ, Pandol SJ, Jeon CY, et al. New-Onset Diabetes, Longitudinal Trends in Metabolic Markers, and Risk of Pancreatic Cancer in a Heterogeneous Population. Clin Gastroenterol Hepatol 2020;18:1812-1821.e7. [29] Lee JH, Kim S-A, Park HY, et al. New-onset diabetes patients need pancreatic cancer screening? J Clin Gastroenterol 2012;46:e58-61. [30] Collins GS, Altman DG. External validation of QDSCORE(®) for predicting the 10-year risk of developing Type 2 diabetes. Diabet Med 2011;28:599–607. [31] Boursi B, Finkelman B, Giantonio BJ, et al. A clinical prediction model to assess risk for pancreatic cancer among patients with prediabetes. Eur J Gastroenterol Hepatol 2022;34:33–8. [32] Boursi B, Finkelman B, Giantonio BJ, et al. A Clinical Prediction Model to Assess Risk for Pancreatic Cancer Among Patients With New-Onset Diabetes. Gastroenterology 2017;152:840- 850.e3. [33] Baecker A, Kim S, Risch HA, et al. Do changes in health reveal the possibility of undiagnosed pancreatic cancer? Development of a risk-prediction model based on healthcare claims data. PLoS One 2019;14:e0218580. [34] Appelbaum L, Cambronero JP, Stevens JP, et al. Development and validation of a pancreatic cancer risk model for the general population using electronic health records: An observational study. Eur J Cancer 2021;143:19–30. [35] Chen W, Butler RK, Lustigova E, et al. Validation of the Enriching New-Onset Diabetes for Pancreatic Cancer Model in a Diverse and Integrated Healthcare Setting. Dig Dis Sci 2021;66:78– 87. [36] Chen W, Zhou B, Jeon CY, et al. Machine learning versus regression for prediction of sporadic pancreatic cancer. Pancreatology 2023;23:396–402. [37] Hamilton W. The CAPER studies: five case-control studies aimed at identifying and quantifying the risk of cancer in symptomatic primary care patients. Br J Cancer 2009;101 Suppl 2:S80-6. [38] Hippisley-Cox J, Coupland C. Development and validation of risk prediction algorithms to estimate future risk of common cancers in men and women: prospective cohort study. BMJ Open 2015;5:e007825. [39] Price S, Spencer A, Medina-Lara A, et al. Availability and use of cancer decision-support tools: a cross-sectional survey of UK primary care. Br J Gen Pract 2019;69:e437–43. [40] Akanuwe JNA, Black S, Owen S, et al. Barriers and facilitators to implementing a cancer risk assessment tool (QCancer) in primary care: a qualitative study. Prim Health Care Res Dev 2021;22:e51. ### Page 22 UK Validation of ENDPAC Page 20 of 23 [41] Chiang PP-C, Glance D, Walker J, et al. Implementing a QCancer risk tool into general practice consultations: an exploratory study using simulated consultations with Australian general practitioners. Br J Cancer 2015;112 Suppl 1:S77-83. [42] Ham C. Money can’t buy you satisfaction. BMJ 2005;330:597–9. [43] Starfield B. Why is the grass greener? BMJ 2005;330:727–9. [44] Thompson L. A Comparative Evaluation of the Structure of Primary Care In the United States and the United Kingdom. 2010. [45] Docteur E, Berenson RA. How does the quality of health care compare internationally? SSRN Electron J 2009. [46] Khan S, Safarudin RF, Kupec JT. Validation of the ENDPAC model: Identifying new-onset diabetics at risk of pancreatic cancer. Pancreatology 2021;21:550–5. [47] Boursi B, Patalon T, Webb M, et al. Validation of the enriching new-onset diabetes for pancreatic cancer model. Pancreas 2022;51:196–9. [48] NHS Digital. National Quality and Outcomes Framework. NHS Digital 2005. https://digital.nhs.uk/data-and-information/publications/statistical/quality-and-outcomes- framework-achievement-prevalence-and-exceptions-data/national-quality-and-outcomes- framework-statistics-for-england-2004-05 (accessed 3 March 2023). [49] McGovern A, Hinton W, Correa A, et al. Real-world evidence studies into treatment adherence, thresholds for intervention and disparities in treatment in people with type 2 diabetes in the UK. BMJ Open 2016;6:e012801. [50] de Lusignan S, Liyanage H, Hobbs R. Health Data Research Innovation Gateway. ORCHID Clinical Informatics Digital Hub 2021. https://web.www.healthdatagateway.org/collection/5626663352808625 (accessed 22 March 2023). [51] Correa A, Hinton W, McGovern A, et al. Royal College of General Practitioners Research and Surveillance Centre (RCGP RSC) sentinel network: a cohort profile. BMJ Open 2016;6:e011092. [52] NHS Digital. Snomed-CT 2017. https://termbrowser.nhs.uk/?perspective=full&conceptId1=363418001&edition=uk- edition&release=v20231122&server=https://termbrowser.nhs.uk/sct-browser- api/snomed&langRefset=999001261000000100,999000691000001104 (accessed 12 December 2023). [53] HbA1c reporting methods set to change. Diabetes UK 2011. https://www.diabetes.org.uk/about_us/news_landing_page/hba1c-reporting-methods-set-to- change (accessed 21 September 2023). [54] Fluss R, Faraggi D, Reiser B. Estimation of the Youden Index and its associated cutoff point. Biom J 2005;47:458–72. [55] Khan S, Chari ST. ENDPAC model for defining a high-risk group for pancreatic cancer: Strengths, limitations and the way forward. Pancreatology 2025;25:1404–6. [56] Rañopa M, Douglas I, van Staa T, et al. The identification of incident cancers in UK primary care databases: a systematic review. Pharmacoepidemiol Drug Saf 2015;24:11–8. [57] Hirst JA, Farmer AJ, Smith MC, et al. Timings for HbA1c testing in people with diabetes are associated with incentive payments: an analysis of UK primary care data. Diabet Med 2019;36:36–43. [58] Mellenthin C, Meier CR, Jick SS, et al. Performance of the EndPAC risk score for pancreatic cancer in new onset diabetes patients with missing data - a database study in the British population. Pancreatology 2025;25:1434–41. [59] Strongman H, Williams R, Bhaskaran K. What are the implications of using individual and combined sources of routinely collected data to identify and characterise incident site-specific cancers? a concordance and validation study using linked English electronic health records data. BMJ Open 2020;10:e037719. ### Page 23 UK Validation of ENDPAC Page 21 of 23 [60] Arhi CS, Bottle A, Burns EM, et al. Comparison of cancer diagnosis recording between the Clinical Practice Research Datalink, Cancer Registry and Hospital Episodes Statistics. Cancer Epidemiol 2018;57:148–57. [61] Badrick E, Renehan I, Renehan AG. Linkage of the UK Clinical Practice Research Datalink with the national cancer registry. Eur J Epidemiol 2019;34:101–2. [62] Margulis AV, Fortuny J, Kaye JA, et al. Validation of Cancer Cases Using Primary Care, Cancer Registry, and Hospitalization Data in the United Kingdom. Epidemiology 2018;29:308–13. [63] Collins GS, Reitsma JB, Altman DG, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): the TRIPOD statement. Ann Intern Med 2015;162:55–63. [64] Van Calster B, McLernon DJ, van Smeden M, et al. Calibration: the Achilles heel of predictive analytics. BMC Med 2019;17:230. [65] Steyerberg EW. Clinical Prediction Models. Springer International Publishing; 2019. [66] Office for National Statistics. Risk factors for pre-diabetes and undiagnosed type 2 diabetes in England - Office for National Statistics 2024. https://www.ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/healthinequaliti es/bulletins/riskfactorsforprediabetesandundiagnosedtype2diabetesinengland/2013to2019 (accessed 15 March 2024). [67] Garg SK, Chari ST. Early detection of pancreatic cancer. Curr Opin Gastroenterol 2020;36:456– 61. [68] Lyratzopoulos G, Abel GA, McPhail S, et al. Measures of promptness of cancer diagnosis in primary care: secondary analysis of national audit data on patients with 18 common and rarer cancers. Br J Cancer 2013;108:686–90. [69] Clift AK, Tan PS, Patone M, et al. Predicting the risk of pancreatic cancer in adults with new- onset diabetes: development and internal-external validation of a clinical risk prediction model. Br J Cancer 2024;130:1969–78. [70] Ke T-M, Lophatananon A, Muir KR. An Integrative Pancreatic Cancer Risk Prediction Model in the UK Biobank. Biomedicines 2023;11:3206. [71] Herrett E, Gallagher AM, Bhaskaran K, et al. Data Resource Profile: Clinical Practice Research Datalink (CPRD). Int J Epidemiol 2015;44:827–36. [72] van Walraven C. A comparison of methods to correct for misclassification bias from administrative database diagnostic codes. Int J Epidemiol 2018;47:605–16. [73] Debray TPA, Vergouwe Y, Koffijberg H, et al. A new framework to enhance the interpretation of external validation studies of clinical prediction models. J Clin Epidemiol 2015;68:279–89. [74] Chowdhury MZI, Turin TC. Variable selection strategies and its importance in clinical prediction modelling. Fam Med Community Health 2020;8:e000262. [75] Holmes JW, Lumens J. Clinical applications of patient-specific models: The case for a simple approach. J Cardiovasc Transl Res 2018;11:71–9. [76] Chari ST, Maitra A, Matrisian LM, et al. Early Detection Initiative: A randomized controlled trial of algorithm-based screening in patients with new onset hyperglycemia and diabetes for early detection of pancreatic ductal adenocarcinoma. Contemp Clin Trials 2022;113:106659. [77] Oldfield L, Stott M, Hanson R, et al. United Kingdom Early Detection Initiative (UK-EDI): protocol for establishing a national multicentre cohort of individuals with new-onset diabetes for early detection of pancreatic cancer. BMJ Open 2022;12:e068010. [78] Claridge H, Cooke EA, Thomas SA, et al. Pancreatic cancer risk assessment tools in primary care: A mixed methods systematic review. J Gastrointest Cancer 2025;56:128. [79] Moriarty AS, Castleton J, McMillan D, et al. The value of clinical prediction models in general practice: A qualitative study exploring the perspectives of people with lived experience of depression and general practitioners. Health Expect 2024;27:e70059. ### Page 24 UK Validation of ENDPAC Page 22 of 23 [80] Wiering B, Lyratzopoulos G, Hamilton W, et al. Concordance with urgent referral guidelines in patients presenting with any of six ‘alarm’ features of possible cancer: a retrospective cohort study using linked primary care records. BMJ Qual Saf 2022;31:579–89. [81] Hamilton W, Mounce L, Abel GA, et al. Protocol for a pragmatic cluster randomised controlled trial assessing the clinical effectiveness and cost-effectiveness of Electronic RIsk-assessment for CAncer for patients in general practice (ERICA). BMJ Open 2023;13:e065232. [82] Hamilton W, Green T, Martins T, et al. Evaluation of risk assessment tools for suspected cancer in general practice: a cohort study. Br J Gen Pract 2013;63:e30-6. [83] Claridge H, Price CA, Ali R, et al. Determining the feasibility of calculating pancreatic cancer risk scores for people with new-onset diabetes in primary care (DEFEND PRIME): study protocol. BMJ Open 2024;14:e079863. [84] Moons KGM, Altman DG, Vergouwe Y, et al. Prognosis and prognostic research: application and impact of prognostic models in clinical practice. BMJ 2009;338:b606. [85] Noble D, Dent T, Greenhalgh T. Time to compare impact and feasibility of prediction models in real life. BMJ 2012;345:e4357; author reply e4360. [86] Dikomitis L, Green T, Macleod U. Embedding electronic decision-support tools for suspected cancer in primary care: a qualitative study of GPs’ experiences. Prim Health Care Res Dev 2015;16:548–55. [87] NICE. Evidence - Suspected cancer: recognition and referral. NICE Guidance 2021. https://www.nice.org.uk/guidance/NG12/Evidence (accessed 8 September 2022). [88] Nicholson BD, Funston G, Williamson C, et al. Is now the time to rethink risk thresholds in cancer referral guidelines? Br J Gen Pract 2025;75:543–4. [89] Bennett P. Pancreatic cancer: The importance of screening in atypical diabetes presentations. Diabetes Prim Care 2018;20:38–41. [90] Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making 2006;26:565–74. [91] Chen W, Zhou B, Luong TQ, et al. Prediction of pancreatic cancer in patients with new onset hyperglycemia: A modified ENDPAC model. Pancreatology 2024;24:1115–22. [92] Cancer Stats Data Hub 2026. https://crukcancerintelligence.shinyapps.io/CancerStatsDataHub/_w_eddb8b4964b74a1bb0a5 b985dca75108/?_inputs_&nav=%22Incidence%20Breakdowns%20and%20Trends%22&app_sel ect_CancerSite=%22Pancreas%22&app_select_Country=%22United%20Kingdom%22 (accessed 20 April 2026). [93] Nelson HD, Pappas M, Cantor A, et al. Harms of breast cancer screening: Systematic review to update the 2009 U.s. preventive Services Task Force recommendation. Ann Intern Med 2016;164:256–67. [94] The Lancet Gastroenterology Hepatology. Pancreatic cancer screening: more harms than benefits? Lancet Gastroenterol Hepatol 2019;4:255. [95] Collins GS, Altman DG. Identifying patients with undetected gastro-oesophageal cancer in primary care: External validation of QCancer® (Gastro-Oesophageal). Eur J Cancer 2013;49:1040–8. [96] Taksler GB, Keating NL, Rothberg MB. Implications of false-positive results for future cancer screenings: Implications of False-Positive Results. Cancer 2018;124:2390–8. [97] Triphuridet N, Yankelevitz DF, Wolf A. A narrative review of lung cancer screening: risks of lung cancer screening. Curr Chall Thorac Surg 2023;5:7–7. [98] Usman Ali M, Miller J, Peirson L, et al. Screening for lung cancer: A systematic review and meta- analysis. Prev Med 2016;89:301–14. [99] Konings ICAW, Sidharta GN, Harinck F, et al. Repeated participation in pancreatic cancer surveillance by high-risk individuals imposes low psychological burden: Low psychological burden of pancreatic cancer surveillance. Psychooncology 2016;25:971–8. ### Page 25 UK Validation of ENDPAC Page 23 of 23 [100] Cipora E, Partyka O, Pajewska M, et al. Treatment costs and social burden of pancreatic cancer. Cancers (Basel) 2023;15:1911. [101] Kaye DR, Min HS, Herrel LA, et al. Costs of cancer care across the disease continuum. Oncologist 2018;23:798–805. [102] Schwartz NRM, Matrisian LM, Shrader E, et al. The potential cost-effectiveness of a risk- based pancreatic cancer screening strategy in new-onset diabetes. J Clin Orthod 2020;38:7017– 7017. [103] Hamilton W. Diagnosing symptomatic cancer in the NHS. BMJ 2015;351:h5311. [104] Møller H, Gildea C, Meechan D, et al. Use of the English urgent referral pathway for suspected cancer and mortality in patients with cancer: cohort study. BMJ 2015;351:h5102. [105] Medina-Lara A, Grigore B, Lewis R, et al. Cancer diagnostic tools to aid decision-making in primary care: mixed-methods systematic reviews and cost-effectiveness analysis. Health Technol Assess 2020;24:1–332. [106] Hart PA, Kamada P, Rabe KG, et al. Weight loss precedes cancer-specific symptoms in pancreatic cancer-associated diabetes mellitus. Pancreas 2011;40:768–72. [107] Price CA, Mold F, de Lusignan S, et al. Sociodemographic profile of people with diagnosed pancreatic cancer in the UK: Retrospective sentinel network cohort study. Stud Health Technol Inform 2025;327:1115–9. [108] Wrench E, O’neil A, Subar D, et al. Mapping pancreatic cancer incidence and risk factors in England between 2013 and 2019. Pancreatology 2024;24:e25. [109] Diabetes UK. 30% of people living with type 2 diabetes in England are undiagnosed, ONS analysis shows. Diabetes UK 2024. https://www.diabetes.org.uk/about-us/news-and-views/30- people-living-type-2-diabetes-england-are-undiagnosed-ons-analysis-shows (accessed 21 November 2025).