Julia Hippisley-Cox and Carol Coupland

Identifying patients with suspected pancreatic
cancer in primary care:
derivation and validation of an algorithm

Abstract INTRODUCTION resonance imaging (MRI) but they need
Pancreatic cancer is the 11th most common better assessment tools to quantify a
Pancreatic cancer has the worst survival for any
cancer in the UK.1 Diagnoses are often made patient’s risk of different types of cancer and
cancer and is often diagnosed late when the late when the cancer is advanced.2 Less than thereby ensure the right patients are sent for
cancer is advanced. Chances of survival are more 20% of patients are suitable for surgery and the right investigations. This would also
likely if patients can be diagnosed earlier.
84% of patients are likely to have died within make efficient use of scare resources.6
Aim a year of diagnosis,1 giving the worst survival Surprisingly, pancreatic cancer is not
To derive and validate an algorithm to estimate
absolute risk of having pancreatic cancer in patients
rate for any cancer.2 However, the chance of mentioned in current guidelines from the
with and without symptoms in primary care. survival is more likely if patients present at National Institute for Health and Clinical
Design and setting an early stage.1 Excellence (NICE) on the referral of patients
Cohort study using data from 375 UK QResearch® There are only a few established risk with suspected cancer.8 This might be
general practices for development and 189 for factors for pancreatic cancer such as age,2 because relatively little is known about the
smoking,2,3 genetic factors,2 chronic aetiology of pancreatic cancer or its
Method pancreatitis,1 and alcohol.1,4 Diabetes may be attendant presenting features. Therefore,
Included patients were aged 30–84 years, free at
baseline from a diagnosis of pancreatic cancer and
a risk factor for pancreatic cancer or an early the pattern of presenting symptoms for
had not had dysphagia, abdominal pain, abdominal manifestation of a growing tumour.5,6 As patients with pancreatic cancer was
distension, appetite loss, or weight loss recorded in there are few established risk factors and investigated with a view to developing an
the preceding 12 months. The primary outcome currently no reliable screening test, it is algorithm to quantify the risk of a patient
was incident diagnosis of pancreatic cancer
recorded in the following 2 years. Risk factors unlikely that there will be a national having pancreatic cancer; this algorithm
examined included: age, body mass index, smoking screening programme for pancreatic cancer; would incorporate both symptoms and
status, alcohol, deprivation, diabetes, pancreatitis, as such, it is likely that most pancreatic baseline risk factors such as age, chronic
previous diagnosis of cancer apart from pancreatic
cancer, dysphagia, abdominal pain, abdominal
cancers will be diagnosed in patients who are pancreatitis, and smoking.
distension, appetite loss, weight loss, diarrhoea, symptomatic and presenting to primary care. QResearch® primary care database was
constipation, tiredness, itching, and anaemia. Cox The challenge is ensuring earlier diagnosis used to develop the risk prediction model as
proportional hazards models were used to develop
to help improve treatment options (for it contains robust data on many of the
separate risk equations in males and females.
Measures of calibration and discrimination example, possibility of surgery) and relevant exposures and outcomes. It is also
assessed performance in the validation cohort. prognosis. Earlier diagnosis could be helped representative of the population where such
Results by increased public awareness of symptoms a model is likely to be used and has been
There were a total of 1415 incident cases of that might indicate pancreatic cancer, such used successfully to develop and validate a
pancreatic cancer from 4.1 million person-years in as weight loss, loss of appetite, and range of prediction models for use in
the derivation cohort. Independent predictors in
both males and females were age, smoking, type 2 abdominal pain.3,6,7 Diagnosis could also be primary care.9–12 Once validated, the models
diabetes, chronic pancreatitis, abdominal pain, improved by more prompt investigation of could be integrated into clinical computer
appetite loss, and weight loss. Abdominal patients who are symptomatic and present to systems to help systematically identify those
distension was a predictor for females only;
dysphagia and constipation were predictors for
their GP.6 In the UK, GPs will soon have better at high risk and alert clinicians to those who
males only. On validation, the algorithms explained direct access to diagnostic investigations might benefit most from further
59% of the variation in females and 62% in males. such as ultrasound, computerised assessment or interventions.9–12 The
The receiver operating characteristic statistics were
tomography (CT) scanning, and magnetic algorithm could also be made available on
0.84 (females) and 0.87 (males). The D statistic was
2.44 (females) and 2.61 (males). The 10% of
patients with the highest predicted risks contained
62% of all pancreatic cancers diagnosed over the
following 2 years. J Hippisley-Cox, MD, FRCGP, MRCP, professor of E-mail:
clinical epidemiology and general practice; Submitted: 21 July 2011; Editor’s response:
Conclusion C Coupland, PhD, associate professor in medical
The algorithm has good discrimination and 4 August 2011; final acceptance: 30 August 2011.
statistics, Division of Primary Care, University of
calibration and could potentially be used to help ©British Journal of General Practice
identify those at highest risk of pancreatic cancer to
Address for correspondence This is the full-length article (published online
facilitate early referral and investigation.
Julia Hippisley-Cox, Division of Primary Care, 13th 27 Dec 2011) of an abridged version published in
Keywords floor, Tower Building, University Park, print. Cite this article as: Br J Gen Pract 2012;
diagnosis; pancreatic cancer; primary care;
Nottingham, NG2 7RD. DOI: 10.3399/bjgp12X616355
qresearch; risk prediction; symptoms.

e38 British Journal of General Practice, January 2012

those with a history • currently consulting GP with first onset of of pancreatic cancer at baseline. when the cancer is advanced and the chance of survival is. therefore. appetite loss. with a recorded red flag symptom13 in the • currently consulting GP with first onset of 12 months prior to the study entry date. September 2010. National Statistics (ONS). both in which was defined as incident diagnosis of terms of discrimination and calibration. drawn from patients registered with • currently consulting GP with first onset of practices between 1 January 2000 and 30 dysphagia (yes/no).6 been using their EMIS (Egton Medical Symptoms that might herald a diagnosis of Information Systems) computer system for pancreatic cancer were also included. as such. loss of appetite (yes/no). the entry date was the date of pancreatic cancer based on a combination first recorded onset within the study period. date of leaving the practice. according to the Office for subsequent 2 years. the general population to help support the A 2-year period was used as this National Early Diagnosis and Awareness represents the period of time during which Initiative. or abdominal which predicts the chances of having distension. All to be recorded in the patient’s electronic practices in England and Wales who had record and that the patient is likely to know. weight loss. well in an independent sample. abdominal distension.13 Patients were censored cancer. This study has developed a new algorithm loss. usually considered a sign. weight patients who are symptomatic by GPs. Clinical outcome definition and diabetes. and would have its own Pancreatic cancer is often diagnosed late pathway for investigation.6 at least a year were included. there needs to be had a minimum of 12 months registration an increased awareness of symptoms prior to study entry. January 2012 e39 . chronic pancreatitis. a red-flag symptom was defined as one that might alarm the patient and • currently consulting GP with first onset of indicate the presence of pancreatic cancer. pancreatic cancer. and those weight loss symptom (yes/no). if a pancreatic cancer during the 2 years threshold equivalent to the top 10% of following the entry date recorded on either patients with the highest risk is selected. The following were • currently consulting GP with first onset of excluded: patients without a postcode.7 which aims to raise public existing cancers are likely to become awareness of the signs and symptoms of clinically manifest. The following predictor derivation dataset and the remaining one. symptoms of dysphagia. or their linked cause pancreatic cancers occurring within the of death record. Study design and data source A prospective cohort study was carried out in Predictor variables a large population of primary care patients Established predictor variables were from an open cohort study using the examined. An open cohort recorded prior to study entry: of patients aged 30–84 years was identified. Two-thirds of Separate analyses were carried out in males practices were randomly allocated to the and females. poor. that is. using the relevant diagnostic codes from the International Classification of Diseases-9 (ICD-9) (157) or the internet as a simple calculator for use by ICD-10 (C25). focusing on those that are likely QResearch database (version 30). in order to the practice. this ensured that all patients make diagnoses earlier. date of death. There Patients entered the cohort on the latest is no reliable screening test so most of the study start date (1 January 2000) and diagnoses are likely to be made in patients 12 months after the patient registered with who are symptomatic. Jaundice was not How this fits in included as this is relatively rare. the patient’s GP record using the relevant UK this will account for 62% of all cases of diagnostic Read Codes. variables were examined. smoking. abdominal pain. abdominal distension (yes/no). loss of • recently consulted a GP with first onset of appetite. related Townsend score.1. this study. using information third to a validation dataset. abdominal pain. For abdominal pain (yes/no). or any of: British Journal of General Practice. The algorithm performed The study outcome was pancreatic cancer. or 2 years after their METHOD study entry date. and encourage those who may have at the earliest date of either diagnosis of symptoms to seek advice earlier. of symptoms and baseline risk factors such as age. For patients with among patients and earlier investigation of incident dysphagia. The sensitivity was high — for example.

4) datasets.1 (14.1) females obtained from the derivation cohort Haemoglobin recorded in preceding year 398 059 (16. Rubin’s rules were used to Non-smoker 1 200 385 (50. which ranges between –6 (most R2 (estimated variation explained by the risk affluent) and +11 (most deprived).0) smoking. Patients are free of a equation in time to pancreatic cancer24).4) 5567 (0. • diabetes (Type1/Type2/no diabetes) at  tiredness in past 12 months (yes/no).4 methods.01 Alcohol intake not recorded 518 328 (21.8) 619 388 (49. main analyses.2) clustering of patients within general Mean BMI (SD) 26.6) 96 443 (7. alcohol intake.4) Current abdominal distension 7985 (0. and smoking validation cohorts.5) combine the results across the imputed Ex-smoker 426 697 (18.3) 16 172 (1.7) Smoking status practices. which were combined with the Prior acute pancreatitis 5029 (0. defined as recorded • smoking status (non smoker.2) 624 352 (50.2) used to model non-linear risk relationships Light smoker (<10/day) 149 044 (6.  constipation in past 12 months • Townsend deprivation score. trivial Derivation and validation of the models The risk-prediction algorithm was [<1 unit/day].5) 96 479 (7.4) the methods implemented in Stata.2) were then applied to the validation cohort Haemoglobin <11 g/dl in preceding year 29 808 (1. Current weight loss 26 239 (1.3) 4929 (0.8) fitted initially and variables were retained if Heavy smoker (≥20/day) 135 113 (5. years (SD) 50.17–20 Five imputations were Characteristic n = 2 364 571 n = 1 243 740 carried out. Baseline characteristics of patients in the derivation and mass index.4) all binary predictor values set to zero.0) 228 970 (18.1) 14 686 (1. a Townsend score is a deprivation score derived from patients’ postcodes. alcohol intake.7) validation cohort to replace missing values Diarrhoea in preceding year 22 377 (0.4) 1 009 931 (81. • age at baseline (continuous. D statistic25 (a measure of discrimination e40 British Journal of General Practice.9) 256 390 (20.8) examined and included in the final models if Moderate or heavy (≥3 units/day) 177 129 (7.6) 8476 (0.80 or >1.22 A full model was Moderate smoker (10–19/day) 180 887 (7.21 Fractional polynomials were Current smoker. study entry. derived from (yes/no).6) 26.2) 2707 (0. BMI = body mass index. Interactions None 512 816 (21.0) 12 688 (1.2) baseline survivor function evaluated at Prior chronic pancreatitis 2208 (0.3) 80 402 (6.8) 627 868 (50.5) they were statistically significant at the 0.8) binary variables) and were statistically Alcohol intake significant at the 0.5) with continuous variables.3 (3. using robust Mean Townsend score (SD)a –0. and Tiredness in preceding year 22 674 (1. The risk equations for males and Itching in preceding year 2615 (0.1) 1206 (0. and • body mass index (continuous).3) 28 578 (2.3) 41 869 (3. to derive absolute risk equations for Prior cancer apart from pancreatic cancer 54 018 (2.4) –0.2 • alcohol status (non-drinker.1) 1454 (0.23 The baseline survivor Symptoms function was estimated based on zero Current dysphagia 15 648 (0. Figures are n (%) unless otherwise specified status and these values were used in the Derivation cohort. amount not recorded 71 668 (3.8) Light (1–2 units/day) 495 561 (21.20 (for Smoking status not recorded 200 777 (8.1) 2 years.0) for body mass index.7) 8507 (0. ranging from • previous diagnosis of cancer apart from 30 to 84) years.01 level. Cox’s proportional hazards Females 1 178 682 (49.0) 50. pancreatic cancer at study entry.3) variable from the final model were used as Type 2 diabetes 78 687 (3.7) 276 449 (22.8) 214 497 (17.4 (4. SD = standard deviation.9) and females separately. Medical history The regression coefficients for each Type 1 diabetes 7269 (0.8) models were used to estimate the Males 1 185 889 (50.8) 129 924 (10.3) and measures of discrimination calculated.0) 39 438 (3.0) 258 963 (20. heavy smoker [≥20 (yes/no) cigarettes/day].4 (4. • anaemia. light haemoglobin <11 g/dl in 12 months (1–9 cigarettes/day). with Current abdominal pain 232 586 (9.5  itching in past 12 months (yes/no).2) between predictor variables and age were Trivial (<1 unit/day) 660 737 (27.2) Multiple imputation was used in the Constipation in preceding year 15 094 (0. • pancreatitis (acute/chronic/none) at study entry.14–16 Multiple imputation was used to replace missing values for body Table 1. using Current appetite loss 10 351 (0. ex. the diagnosis of pancreatic cancer at baseline.0) they had a hazard ratio of <0. moderate (10–19 before study entry or the 60 days after cigarettes/day).9) 12 233 (1.3) 3986 (0.3) 2 years of follow-up. light [1–2 units/day].7) values of centred continuous variables.6) level. Validation cohort.2) coefficients for each risk factor for males Mean age.9–12.4) weights.6) variance estimates to allow for the BMI recorded prior to study entry 1 877 243 (79.7) 74 140 (6.5) 93 705 (7.9) 358 233 (28. January 2012 .1 (15. developed and validated using established moderate/heavy [≥3 units/day]). patients’ postcodes (continuous).2 (3.  diarrhoea in past 12 months (yes/no).

01%) patients with a history of pancreatic cancer were also excluded. of 30–39 40–49 50–59 60–69 70–79 80–84 30–39 40–49 50–59 60–69 70–79 80–84 which 375 were randomly assigned to the Age group (years) derivation dataset. a total of 1415 incident all other terms in the table and for age. and body pancreatic cancer for males and females in the derivation cohort mass index (not shown.84) 5.35 to 4.46) 3. A total of Abdominal Abdominal pain 2 538 615 patients aged 30–84 years were distension identified in the derivation cohort.37 (1.46 to 4.27 (2. loss. All available data on the database were used to maximise the power and 2000 generalisability of the results. restricting the Females Males analyses to patients who had the outcome 3000 within 2 years or had at least 2 years of Incidence rate per 100 000 follow-up. 5%.84 to 19. and Appetite loss Dysphagia 124 458 (4.9%) patients without a recorded Weight loss Townsend deprivation score were excluded.74 to 3.9)b females and increased steeply with age.94 (1.53 (1. appetite higher values indicate better discrimination). Adjusted hazard ratios for Adjusted hazard ratios females (95% CI) for males (95% CI) Incidence rates for red-flag symptoms Non-smoker 1 1 Overall.91 (1. of these.09 (3. characteristic (ROC) curve (ROC curve A total of 1 342 329 patients aged statistic) at 2 years were calculated. 26 239 with weight Heavy smoker 2.66 to 2.36 to 2.25) 1.5%. dysphagia.03) patients with dysphagia were identified.93 to 8.07 (1. The excluded. using these thresholds.46 (1. the 12 months prior to entry to the study. and positive and 1.04 to 2. and 10% of patients at highest estimated The baseline characteristics of each risk of pancreatic cancer at 2 years. specificity. in the derivation cohort.32 (1.23) weight loss were similar in males and Current weight lossa 3. Current abdominal paina 4.20 to 2. 1%.1%.90 (2. Moderate smoker 2. with the remainder assigned to a validation cohort.20) 10 351 with appetite loss. as well as 96 (0. Hazard ratios adjusted for year follow-up period.43 to 4. 1 243 740 patients for analysis.01) symptom.10) 1. cohort were very similar. alcohol intake.15 (1.61) loss. 1000 RESULTS Overall study population 0 Overall.5 (7.11) Abdominal pain was more common in Current abdominal distensiona 3. suitable for use with survival data where at least one red-flag symptom recorded in abdominal pain. which were age-2 and age3 for females and age-1 for males. 232 586 with abdominal pain. 161 (0. 15 648 Ex-smoker 0. Stata (version 11) was used for all analyses.61 to 5. leaving the thresholds for the 0.01%) with a history observed risk was obtained using the of pancreatic cancer.12 to 1.39 (1.71) a Compared with person without this characteristic. in males and females in the derivation cohort. Constipation in last yeara NS 1. during the 2- model for males also included an interaction between weight loss and the age term. as shown in Table Sensitivity. As in another study.17 to 8.88 (1.3%) patients mean predicted risks at 2 years with the without a recorded Townsend score were observed risk by tenth of predicted risk. as well as a further 49 425 (1. it tended to decrease with age in Current dysphagiaa NS 2. 30–84 years were identified in the validation Calibration was assessed by comparing the cohort. 0.67) Light smoker 1. The incidence rates for abdominal Current symptoms and symptoms in preceding year distension.11 (1.60 to 4.23) 1. and 27 646 (2. appetite loss.48) 1.63 (1. at mean age in males.48 to 6. available from the authors).97 (0.50) NS females.44 (1.23 (4.10) females but increase with age in males. cases of pancreatic cancer arising from British Journal of General Practice. Figure 1 shows Type 2 diabetes 2. and 7985 Medical history with abdominal distension.65 to 3. NS = not significant. and Current appetite lossa 3. and weight loss per 100 000 person years and the area under the receiver operating leaving 2 364 571 patients for analysis.52) the age–sex incidence rates of each Chronic pancreatitis 3. 564 QResearch practices in England and Wales met the inclusion criteria.10 the patterns of negative predictive values were calculated missing data supported the use of multiple imputation to replace missing values for Table 2. Adjusted hazard ratios (95% CI) for the final model for smoking status. bInteraction term.82) 2.76 to 2.9%) patients with Figure 1. Incidence rates of dsyphagia.03 to 2.04 (1. January 2012 e41 .56) 12.58) 2.77 to 1. abdominal distension. The Overall in the derivation cohort.35 to 2.56 (1.1%) with Kaplan–Meier estimate evaluated at 2 years. 70 847 (5.68 to 5. The models also Incidence rates of pancreatic cancer included fractional polynomial terms for age. at least one red-flag symptom recorded in The validation cohort was used to define the 12 months prior to study entry.

6%. dysphagia and constipation such as abdominal ultrasound.60) elevated with appetite loss (2. Risk also increased integrated in to GP computer systems to with the amount smoked. Alternatively.9) was less steep (Figure 2. Adjusted hazard ratios for pancreatic There were 612 cases (78.45 to 2. loss.7%) Figure 4 shows the mean predicted 10 identified from the linked death record.6% of cases diagnosis of pancreatic cancer — higher values females.85 to 0. 781 incident cases well calibrated.4-fold for heavy smokers and cancer based on information already 1. using a D statisticb 2.88) interaction between weight loss and age in ranged from 1. cancer within this group out of 781 new cohort The magnitudes of the hazard ratios in cases identified in the validation cohort.1-fold higher). in order 30 35 40 45 50 55 60 65 70 75 80 85 tended to be higher in males than females to assess the calibration of the model in the Age (years) (data not shown).87 (0. The risks recorded in the patient’s electronic health were also elevated in females with type 2 record.1 to 64. tenth for males and females.84 for females and 0. appetite loss (3. that the risk-prediction equations explained This research has developed and validated a 59% of the variation in time to diagnosis in new algorithm designed to quantify the e42 British Journal of General Practice. 10 males and females were age. weight loss.61 (2. males were generally similar to those for which accounted for 62. 4 149 461 person years of observation were females and 62% of the variation in males.2-fold higher).9%. abdominal distension was not a of pancreatic cancer had a diagnosis of type indicate more variation is explained. appetite loss. and weight loss (3. Software significantly associated with increasing age implementing the algorithm could be as shown in Figure 2. The sensitivity of D statisticb 2. January 2012 . identified. pain (4.9-fold with a 2-year risk score of >0. chronic pancreatitis. with a small Figure 2. Unlike in to 39. abdominal assessment.44 for females and 2. weight loss. validation cohort. Box 1.0-fold higher). The results could help inform the Weight loss Abdominal distension was a predictor for decision to undertake further investigations females only.4% of all new cases Mean (95% CI) females. Table 3.27 to 2. or appetite 0 status. The Females relative increase in risk with increasing age positive predictive value (PPV) with this R2 statistica (%) 58. as shown in Figure 3. 18. MRI. smoking abdominal pain.44 (2. 20 GP record and an additional 335 (23.4 to 61. Validation statistics higher). a measure of discrimination — higher values indicate better discrimination. There was close of pancreatic cancer were identified arising correspondence between predicted and Hazard ratios compared with age = 45 years.86) Males abdominal pain (5. and constipation (1. calculate the risk of a patient having an compared with non-smokers.8% for abdominal pain.5-fold for light smokers (Table 2). type 2 diabetes. There was also a significant an approach based on single symptoms ROC statisticc 0. using the GP record and an additional 169 tenth in males. or endoscopic retrograde variables examined were not independent cholangiopancreatography.9-fold 45. the risks were existing but as yet undiagnosed pancreatic increased by 2. giving a crude rate of 34 cases per The D statistic was 2. Overall.1-fold higher). ROC statisticc 0.3%) had a sensitivity of R2 statistica (%) 62.7 (55. The risks were also threshold was 0. (21.2% for abdominal distension R statistic shows explained variation in time to a 2 males. 2 diabetes at baseline.61 Adjusted hazard ratio 50 40 100 000 person years.84 (0.6 %) from the linked death record. The area under the ROC statistics 30 cases (76. but the of pancreatic cancer (sensitivity).8) (2. dysphagia a 2-year risk score >0. threshold based on the top 5% of risk (that is. observed 2-year risks within each model giving a rate of 36 per 100 000 person years.3 % of 1415) identified using the were 0.87 for males. abdominal distension The 90th centile defined a high-risk group (3.4%) identified degree of over-prediction in the highest cancer by age in males and females. For example. CT Hazard ratios compared with age = 45 years. models.77) higher).3-fold higher) There were 487 new cases of pancreatic for the risk prediction algorithm in the validation (Table 2). The other scan. were predictors for males only.5-fold higher). as shown in Table 2). The scores and the observed risks at 2 years 0 incidence rates increased with age and within each tenth of predicted risk. bD statistic is significant predictor in males. Figure 3.6-fold higher). Some clinical examples are shown in 30 35 40 45 50 55 60 65 70 75 80 85 Age (years) abdominal pain. Patients at highest risk could be diabetes (2. chronic identified and recalled for a clinical pancreatitis (3. There were 1080 for males.2% (Table 4). Independent predictors in both an alarm symptom such as dysphagia.82 to 0. systematic risk stratification for a population Risk of pancreatic cancer in females was of patients aged 30–84 years. Adjusted hazard ratio for weight loss by risk factors so were not included in the final The algorithm could also be used for age in males. from 2 184 336 person years of observation.0 (59.2-fold higher). the model was Females Males In the validation cohort. cROC statistic is a Validation DISCUSSION measure of discrimination — higher values The validation statistics (Table 3) showed Summary indicate better discrimination.3% and a PPV of 0. Individual risk assessment and thresholds Adjusted hazard ratio 50 Predictor variables One potential use for this algorithm is within 40 Table 2 shows the predictor variables consultations with individual patients 30 selected for the final models for females and particularly if they present with new onset of 20 males.

9 0.9 Current symptoms Abdominal pain n/a 877 133 470 93 792 311 39. appetite loss.9 Top 0.9 Dysphagia n/a 965 494 770 5431 11 1. 0.4 99.7 2.9 100. recall. of cases. from the UK ONS and the study is therefore pancreatic cancer over 2 years by tenth of predicted which is either currently present or likely to likely to have picked up the majority of cases risk applying the risk prediction scores to the become manifest within 2 years.0 Top 1% risk 0.9 964 373 685 6552 96 12. representativeness.9 0. The reflects the improved ascertainment algorithm is based on simple clinical resulting from including of diagnoses variables which can be ascertained in recorded in either the primary care or the clinical practice. and followed-up. British Journal of General Practice.00 bias.3 99.3 0. loss of appetite. it information recorded by primary care Table 4. and associated symptoms such as weight loss. UK general practices have good levels of accuracy and completeness in recording Comparison with existing literature clinical diagnoses.3 95.3 931 077 427 39 848 354 45.7 0.4 99.9 The positive predictive values (PPVs) are an average for patients in each category.9 Abdominal distension n/a 967 478 772 3447 9 1.4 99. type 2 diabetes. the GP and then entered into the patient’s and lack of selection. The algorithm performed cause of death record. hospital discharge letters which are sent to duration of follow-up. The database has linked cause of death Figure 4.20 diabetes.1 0.9 4.40 treated.4 0. chronic estimates suggesting good ascertainment pancreatitis. dCriterion met does have disease. linked cause of death data.1 99.3 99. this is the first ascertainment bias. n/a = not applicable.2 99. Specificity.9 Appetite loss n/a 967 570 754 3355 27 3. The incidence rate in algorithm of its kind and is based on age. and abdominal pain.8 90. aCriterion not met does not have disease.1% risk 2. Mean predicted risk and observed risk of absolute risk of having pancreatic cancer.30 significance of established risk factors such 2-year risk (%) as age.2 99.5 99.6 96.3 1. and constipation.9 Chronic pancreatitis n/a 970 094 777 831 4 0.9 Top 0.3 99. PPVs for individuals can be calculated using the web calculator to take their characteristics into account.4 91. Patients diagnosed with pancreatic cancer in hospital Strengths and limitations will have the information recorded in Key strengths of the study include size. chronic pancreatitis.10 Limitations of the study include lack of formally-adjudicated outcomes. Not all patients with 0 5 10 0 5 10 symptoms will attend their GP. this study’s population is close to European smoking status. type 2 0. abdominal UK cancer registry reports which probably distension.27 It is higher than rates reported in weight loss. Patients who die of well in a separate validation sample with pancreatic cancer will be included on the good discrimination and calibration.5 99.6 100. has been conducted in the setting where the Female Male majority of patients in the UK are assessed. % Type 2 diabetes n/a 940 800 636 30 125 145 18. predictive predictive Criteria threshold % negativea negativeb positivec positived % % value. and responder electronic record. information 0. To the of pancreatic cancer. Comparison of strategies to identify patients at risk of having a diagnosis of pancreatic cancer in the next 2 years based on the validation cohort Positive Negative Risk True False False True Sensitivity.6 0. bCriterion not met does have disease. bias.8 970 423 758 502 23 2.9 0.3 967 852 716 3073 65 8.0 Top 5% risk 0. abdominal pain.5 99. and missing data. authors’ knowledge.6 99.2 884 670 294 86 255 487 62.5 99. not all symptoms will be reported or Observed risk Predicted risk recorded. and in those Tenth of predicted risk who do. cCriterion met does not have disease.3 99. dysphagia. smoking.26 The authors consider While the study was reliant on accuracy of that the study has good face validity since.5% risk 1.9 Risk threshold Top 10% risk 0.8 99.8 99.9 Weight loss n/a 961 571 720 9354 61 7.0 0. January 2012 e43 . The algorithm has good face validity as it confirms the 0. % value. alcohol status. 0.9 99. thereby minimising validation cohort.

algorithm can identify 10% of the population depression. specificity. error cannot be shows the possible thresholds along with around a risk threshold. such as vomiting. who is a heavy smoker and presents with abdominal inform subsequent cost-effectiveness pain. could be also be used by members of the results suggest that investigation of the Acknowledgements pancreas should be considered alongside public via a simple web calculator which We acknowledge the contribution of EMIS could then prompt symptomatic patients to other upper gastrointestinal investigations practices who contribute to QResearch® and consult their GP. Carol Coupland is where curative treatment is more likely to be predictive on multivariate analysis. loss of appetite. Given the associate professor of medical statistics at possible. Clinical examples negative predictive values. January 2012 . and constipation the estimated risk would be 3. the algorithm There was no external funding.4%. modelling and the choice of thresholds If he also has a history of chronic pancreatitis. Following physicians.28 Currently however. Provenance data sources. Open Source Software under the GNU lesser systematically assessed. and inform the oesophagus resulting in dysphagia. Julia Hippisley-Cox is also a paid NICE guidelines on suspected cancer fail to mention pancreatic cancer despite its traditionally thought to be associated with director of ClinRisk Ltd. EMIS for expertise in establishing.2% estimated risk of having pancreatic cancer. Table 4 performance or results in reclassification the nausea. and dysphagia has a 10.9%. Competing interests registries in the near future which is likely to although there were only one-tenth of the Julia Hippisley-Cox is professor of clinical increase the accuracy and completeness of number of patients with type 1 diabetes epidemiology at the University of Nottingham this information. and implementation of clinical risk algorithms constipation in males) were also identified within clinical computer systems to help this symptom based tool can lead to earlier identification of pancreatic cancer at a stage but which remained independently improve patient care. database regarding the precise type of pancreatic cancer it was found that Web calculator cancer which means it was not possible to pancreatic cancer is twice as likely to occur Here is a simple web calculator to implement include the precise type of pancreatic cancer in patients with type 2 diabetes than those the QCancer® (pancreatic) algorithm. there is While the study was not designed to limited information on the QResearch determine whether type 2 diabetes causes Freely submitted. These could be explored in Discuss this article in which approximately 62% of all new future versions of the algorithm to Contribute and read comments about pancreatic cancer cases are likely to be determine whether they improve its this article on the Discussion Forum: diagnosed over the next 2 years. which produces pancreatic cancer (such as. validated similar outcomes and exposures scope of the present study). abdominal The algorithms have been developed in distension. compared with type 2 diabetes so a type 2 and co-director of QResearch® — a not-for. who is a heavy smoker with type 2 diabetes and no symptoms has a 0. backache. which in the outcome (that is. If the same male had be considered for preventative measures. For example. The QResearch database will increased risk of pancreatic cancer among http://www. all patients with diabetes is low (0. although the absolute risk across is publically available alongside the paper endocrine and exocrine tumours). location of the pancreas.1%. be linked with information from cancer patients with type 1 diabetes was found. which is outside the scope of this paper. developing integrated into GP clinical computer There may be other symptoms which could and supporting the database. dyspepsia. externally peer reviewed. the sensitivity.3%). using questionnaire data and found levels of could potentially be used in clinical practice Ethics committee completeness and accuracy in similar GP to identify those at highest risk of having All QResearch® studies are independently databases to be good. abdominal pain. one pancreatic cancer to facilitate early referral reviewed in accordance with the QResearch® systematic review reported that on average and investigation and so help earlier agreement with Trent Multi-Centre Ethics 89% of diagnoses recorded on the GP identification of patients with pancreatic Committee (UK). electronic record are confirmed from other cancer. one cohort and validated in a separate • A 50-year-old male. the quality of information is likely independent external validation and cost- Funding to be good since previous studies have effectiveness modelling (which is outside the This study was undertaken by ClinRisk Ltd. Further research is needed to assess whether use of distension in females. e44 British Journal of General Practice. the estimated risk would be 44. positive and Box 1. • A 75-year-old female. The algorithm could be for patients presenting with dysphagia. and loss of appetite.3% estimated risk of having pancreatic cancer.29 For example. distinguish between without it. This GPs at the point of care to assess risk in fundus of the stomach or lower end of the work and any views expressed within it are symptomatic patients. the estimated risk would be 34. this study’s affiliated bodies or organisations. dysphagia. the algorithm confirmed by other studies. appetite loss. abdominal software to ensure the reliable and updatable substantial morbidity and mortality. If she also has weight loss. Further research profit organisation which is a joint Implications for research and practice into the potential mechanisms underlying partnership between the University of This study could help raise awareness of this association for type 2 diabetes may be Nottingham and EMIS (leading commercial symptom complexes predictive of pancreatic cancer especially since current warranted. The algorithms systems and used to generate a list of high predict pancreatic cancer that were not presented in this paper will be released as risk patients who could then be recalled and included in this study. has an 11. who is a moderate smoker and has abdominal pain. supplier of IT for 60% of general practices in Several additional symptoms not the UK). The algorithm performed well with good discrimination and calibration. fever.rcgp.1% cohort representative of the patients likely to estimated risk of developing pancreatic cancer in the next 2 years. This is intended to • A 70-year-old male with type 2 diabetes. No and open source software or the and GPL v3.28. dysphagia. the stage. it is possible that a the University of Nottingham and a paid Although primarily designed to be used by tumour could result in pressure on the consultant statistician for ClinRisk Ltd.qcancer. If solely those of the co-authors and not of any decision to investigate or refer.

Dhingsa R. 340: c2442. 363(9414): 1049–1057. The National Awareness and Early Diagnosis Initiative in 21. Predicting the risk of Chronic Kidney Disease in 1991. Khan NF. BMC Fam Pract 2010. Research Database. 10. et al. Diabet Med 2011. 14. Royston P. Harrison SE. UK CR. J Clin Epidemiol 2006. An independent and external validation of QRISK2 2010. Br J Clin Pharmacol 2010. Robson J. risk in England and Wales: prospective derivation and validation of QRISK2. QRISK cardiovascular risk prediction: a prospective open cohort study. Ambler G. 59(10): 1092–1101. Alcohol 2005. Lemeshow S. Benjamin IS. van Veen M. Lancet 2004. diagnosis of pancreatic cancer. Charlton J. Palaniappan P. New York. Applied logistic regression. London: National Institute for 22. Validation and validity of diagnosis of cancer in primary care: cohort study using General Practice diagnoses in the General Practice Research Database: a systematic review. BMJ 2010. Schoonen WM. Missing data: our view of the state of the art. 1989. Stat Med 2004. Jick SS. 60(9): 979. 25. Ann Oncol 2007. Ferlay J. introduction and summary of the symposium. Royston P. BMJ 2001. 60(572): e128–136. New York. Hippisley-Cox J. Ann Oncol 1999. J Clin Epidemiol 2007. Khalsa J. Multiple imputation for non-response in surveys. 323(7326): 1413–1416. An independent external validation and evaluation of 2011. Harrell FJ. January 2012 e45 . Cancer of the stomach and pancreas. 334(7602): 1040. 28(5): 599–607. NY: John BMJ 2008. Coupland C. 101 (Suppl 2): S1–4. BMJ 2. Altman DG. Predicting cardiovascular 964–974. Imputation is beneficial for handling missing data 5. Mechanisms of alcohol-associated cancers: Academic Medicine Group. Predicting risk of osteoporotic fracture in men and women in England and Wales: prospective derivation and validation of 24. 28(5): 9. Schafer J. Richards MA. Wolff R. Men and Women in England and Wales: prospective derivation and external 27. 8. Moons KGM. Pancreatic cancer statistics — key facts. Rose PW. NICE. Br J Gen Pract 15. Graham J. General Practice Research Database: a systematic review. Hosmer D. Int J Epidemiol 1999. Coupland C. Stijnen T. Coupland C. 11: 49. Referral guidelines for suspected cancer. 19. Diabetes and the risk of pancreatic cancer. Altman DG. No authors listed. Collins GS. London: Cancer Research. 17. Boniol M. Explained variation for survival models. 16. 7. BMJ 2004. 18(3): 581–592. Latinovic R. Wiley & Sons. BMJ 2009. External validation of the QDScore for predicting the 29. Hippisley-Cox J. Alarm symptoms in early 28. Coupland C. Hippisley-Cox J. Collins GS. Abbruzzese JL. 13. Gulliford MC. 35(3): 155–160. Jones R. et al. diabetes in England and Wales: prospective derivation and validation of 26.10 (Suppl in predictive models. 1. Pancreatic cancer. Recent developments in imputation of missing predictor values was preferred. Jick H. Thomas SL.REFERENCES cardiovascular disease risk score: a prospective open cohort study. Serrano J. Academic medicine: problems and solutions. Takhar AS. et al. Gullo L. Methods 2002. Vinogradova Y. 7(2): 147–177. Br J Cancer 2009. 23(5): 723–748. Inc. The use of fractional polynomials to model Health and Clinical Excellence. 23. Sauerbrei W. 302(6779): 766–768. BMJ 2009. BMJ 2007. continuous risk variables in epidemiology. BMJ 12. Royston P. 298(6673): 573–579. Derby LE. QFractureScores. Donders RART. Validity of diagnostic coding within the 10-year risk of developing Type 2 diabetes. John Wiley. 338: b880. 1987. Psychol 3. A new measure of prognostic separation in survival 11. Bowles MJ. Li D. 69(1): 4–14. 339: b2584. Autier P. Altman DG. Xie K. Hippisley-Cox J. NY: England: assembling the evidence. 329(7467): 668–673. 2005. Predicting risk of type 2 data. Estimates of the cancer incidence and validation of the QKidney(R) Scores. 339: b4229. practitioner based computerised data resource in the United Kingdom. mortality in Europe in 2006. The 4. 336(7659): 1475–1482. Using the outcome for 6. 6: 1–14. Lobo DN. British Journal of General Practice. Rubin DB. Herrett E. Stata J 2006. 4): 79–81. Sauerbrei W. Steyerberg EW. et al. BMJ 1989. Collins GS. Validation of information recorded on general QDScore. 18. 2009. Purohit V. 20.