Detectable Symptomatology Preceding the Diagnosis of Pancreatic Cancer and Absolute Risk of Pancreatic Cancer Diagnosis

Abstract

The survival duration for pancreatic cancer is short. Given its low lifetime risk (1.5%), established factors for the disease have insufficient specificity to identify individuals at high risk of nonfamilial cancer, and prediagnostic signs and symptoms are vague and not limited to pancreatic causes. We considered whether statistical models that incorporated both risk factors and prediagnosis symptomatology could improve prediction enough to provide practical risk estimates. We combined US Surveillance Epidemiology and End Results (SEER) incidence data from 2008 to 2010 with regression models from representative case-control data from Connecticut (2005–2009) to estimate age- and sex-specific 5-year absolute risks of pancreatic cancer diagnosis. Our risk model included current cigarette smoking (adjusted odds ratio (OR) = 3.3, 95% confidence interval (CI): 2.1, 5.0), current use of proton pump-inhibitor antiheartburn medications (OR = 6.2, 95% CI: 1.7, 23), recent diagnosis of diabetes mellitus (OR = 4.8, 95% CI: 2.2, 11), recent diagnosis of pancreatitis (OR = 19, 95% CI: 3.1, 120), Jewish ancestry (OR = 1.8, 95% CI: 1.1, 3.1), and ABO blood group other than O (OR = 1.3, 95% CI: 1.0, 1.8). In total, 0.87% of controls with combinations of these factors had estimated 5-year absolute risks greater than 5%, and for some, the risks reached more than 10%. Combining risk factors for pancreatic cancer with detectable prediagnostic symptomatology can allow investigators to begin to identify small segments of the population with risks sufficiently high enough to make screening efforts among them potentially useful.

Editor's note:An invited commentary on this article appears on page 35, and the authors’ response appears on page 39.

After lung, breast or prostate, and colorectal cancer, carcinoma of the pancreas is the fourth most frequent cause of cancer death in the United States, with the occurrence in women now parallel to that in men. In 2015, 49,000 Americans will be diagnosed with pancreatic cancer, and more than 40,000 will succumb to it (1). In the United States, an individual's probability of ever developing pancreatic cancer is 1.5% (1). Survival with pancreatic cancer is meager. On a population basis, the 1-year relative survival fraction is approximately 15% and 5-year survival less than 2% (2). The majority of patients present with advanced disease; however, surgical resection in early stages can extend survival by a year or more in what is otherwise a progressive and fatal course (2).

Known etiologic factors for pancreatic cancer explain perhaps 30%–40% of the disease. Germline mutations in the breast cancer 2, early onset gene (BRCA2), cyclin-dependent kinase inhibitor 2A, multiple tumor suppressor 1 gene (CDKN2A), and other genes account for 5%–10% of cases (3), and ABO blood group variation accounts for approximately 17% (4). Long-standing adult-onset diabetes mellitus is responsible for approximately 7% of cases (5). Apart from these causes, the one established risk factor for pancreatic cancer is cigarette smoking. Current smokers have at least double the risk of nonsmokers, with trends toward increasing risk according to frequency or duration of smoking and decreasing risk with greater time since cessation (6).

The development of pancreatic cancer is typically accompanied by physical signs and symptoms with patterns that can change during the years and months leading up to diagnosis. Unintentional weight loss, epigastric or back pain, anorexia, nausea or early satiety, subjective epigastric bloating, jaundice, heartburn, pruritus, dysgeusia, and other signs and symptoms have been reported, with weight loss, pain, and bloating the most common. Symptoms generally occur within 6 months of diagnosis (7–11). The nonspecific nature of these complaints limits their clinical utility in directing attention toward a pancreatic neoplasm, and their close temporal proximity to the cancer diagnosis reduces possible benefits of early detection. In addition, adult-onset secondary diabetes mellitus and acute pancreatitis can be induced by pancreatic cancer or by its adjacent pathophysiology as early signs of the disease. However, pancreatic cancers accompanying new diagnoses of diabetes mellitus or episodes of pancreatitis are relatively infrequent. Serum markers for pancreatic cancer among persons with diabetes have not yet identified usable diagnostic factors (12).

General population screening for pancreatic cancer is ineffective because the disease has such a low population risk (1.5% over the lifetime); thus, the probability of developing it during 5 or 10 years at midlife is negligible, and only after age 60 years does the bulk of the lifetime risk begin to accumulate. At such low frequencies, imperfect specificities of even very good risk-factor screening algorithms or laboratory tests produce many more false-positive results than true positives (13). Recent consortial work has shown that simple models of presently known risk factors for pancreatic cancer do not achieve sufficiently large relative risks for combined exposures to surmount the specificity limitation (5). Even among adults with new-onset diabetes, the fraction with undiagnosed pancreatic cancer is so low that using postprandial pancreatic polypeptide levels as a screening test is ineffective (12).

In this context, we considered whether statistical models that incorporated both risk factors for pancreatic cancer and concrete symptomatology detectable before diagnosis could improve risk prediction enough to provide practical estimates of risk. If so, such models could be useful in identifying individuals with risks high enough to make diagnostic screening modestly practical and efficient and could help to increase the fraction of patients with resectable tumors and improved chances of survival. We examined such models in our large pancreatic cancer case-control study in Connecticut, along with recent US Surveillance Epidemiology and End Results (SEER) incidence data, to calculate absolute risks of the disease.

METHODS

Study subjects and biospecimens

Our case-control study of pancreatic cancer in Connecticut has been described in detail previously (14). In total, 1,081 potentially eligible individuals 35–83 years of age with newly diagnosed pancreatic cancers were identified in Connecticut hospitals between January 2005 and June 2009. Physicians approved 83% of subjects for participation in the study. In-person interviews were thus sought for 895 patients and completed for 403 (45%): 50 were unable to be contacted or did not have landline telephones, 333 had died before study contact or were too ill to participate, and 109 refused to participate. Eligibility was confirmed through examination of clinical or pathology records. Over the same time period, we used random digit dialing (preceded by letters sent to eligible households) to identify potential control subjects. Selected landline telephone numbers were first searched through reverse-directory lookup to find addresses in order to mail study letters before telephone contact for subject eligibility ascertainment. Control subjects were frequency matched to cases by age group and sex. We identified 1,137 potentially eligible control subjects and interviewed 715 (63%) in person. We were unable to contact 140 subjects, and 282 refused to participate. At interview, case and control subjects provided signed informed consent, after which questionnaire interview was conducted and venipuncture blood samples (93.3%) or saliva samples (5.6%) were obtained. Approval for the study was received from the human subjects institutional review boards of Yale University and the State of Connecticut Department of Public Health, as well as by the 30 Connecticut hospitals from which cases were identified.

Within 4 hours of interview, biospecimens on freeze packs were transported to our laboratory. Erythrocytes, plasma, and buffy coat were separated, aliquotted, and stored at −80°C until analyses. Samples were tested for Helicobacter pylori and cytotoxin-associated gene A (CagA) seropositivity by using commercial immunoassay kits (14). ABO blood group was determined by using standard erythrocyte antiserum agglutination testing on the fresh blood samples (14). In the present analyses, we included information from all 362 case and 690 control subjects who provided blood specimens.

Study data and statistical analyses

Our study questionnaire was used to obtain detailed information on cigarette smoking and other tobacco use, alcoholic beverage consumption, usage of medications for relief of heartburn or acid reflux, usual adult height and weight, family history of pancreatic cancer, Jewish ancestry, and diagnoses of pancreatitis and diabetes mellitus. Subjects were asked whether they were currently smoking cigarettes or when they had quit, when the pancreatitis or diabetes had been diagnosed, and for each heartburn medication, when they had started using it and whether they were still doing so or when they had stopped. At interview, the names of both prescription and over-the-counter heartburn medications were presented to subjects on a card for recognition. In analyses, these medications were grouped into proton-pump inhibitors, histamine-receptor antagonists, and antacids; other types were used infrequently. We did not inquire about the timing of current versus past use of alcoholic beverages.

Unconditional logistic regression methods were used as a basic framework to estimate odds ratios. The GLIM program was used to carry out the regression computations (15). All P values are 2-sided. Analyses were adjusted for age at interview, sex, race (white vs. other), reported history of pancreatic cancer in first-degree relatives, H. pylori seropositivity, and CagA seropositivity. In the final models, we did not adjust for body mass index because it had no significant associations in our data and no effects on other model factors, or for alcohol intake because no timeframe of consumption was available. We considered models using both categorized and continuous covariates. Most terms in the models were included in the standard fashion as indicator or linear covariates; however, for heartburn medication usage, cigarette smoking practices, and histories of diabetes mellitus and pancreatitis, we were interested in temporal inverse-duration aspects before diagnosis (12) and so modeled exponential decay of the log odds ratio according to years in the past when the medications were started, number of years since smoking cessation, number of years in the past when diabetes was diagnosed, and number of years in the past when pancreatitis was diagnosed. Each of these 4 factors was thus included in regression models as an interaction term in the form β₁I exp(−β₂t), where I was an indicator variable (0 or 1) denoting personal history of the particular factor and t was its time-in-the-past variable (in years). All of these interaction terms showed continuous log odds ratio decays that were consistent with models in which the odds ratios were stratified according to categories of years in the past. These continuous nonlinear interaction-decay models were fitted in GLIM using the iterative linearization method (16). Confidence limits for the fitted odds ratios as functions of time t before diagnosis were calculated on the log scale using the delta method according to estimated $\text{standard}\text{error}=\{\text{var}({\beta }_1)+{\beta }_1^2{t}^2\text{var}({\beta }_2)-2{\beta }_{1}t\text{cov}({\beta }_1,{\beta }_2){\}}^{1/2}\text{exp}(-{\beta }_{2}t)$ and then exponentiated.

To estimate the absolute risk of developing pancreatic cancer over 5-year periods of follow-up according to age, sex, and various risk factors, we started with recent SEER data (17) calculated by the DevCan program (18). These data give the average sex-specific probabilities of developing cancer in a specific age range for populations covered by the SEER registries, including the state of Connecticut. Given that the controls in our case-control study comprise a representative sample of Connecticut adults, with sampling frequencies adjusted only according to age and sex, we estimated the average relative risk in our controls according to their modeled covariates by calculating a weighted average,

$$\text{Average}\text{relative}\text{risk}=\frac{1}{n}\sum _i{w}_i{\text{e}}^{{\sum }_j{c}_j{\beta }_j{x}_{ij}},$$

where the sum on i is over all n controls, the weights w_i represent the relative frequency of controls of the sex and age group (5-year increments) in the SEER population data compared with the fraction of controls of that sex and age group in our study $(\sum w_i=1),$ and the covariates x_ij and their parameter estimates β_j represent the factors of interest for estimating absolute risk from our adjusted regression model. When estimating average relative risk, only the terms of interest and not all of the model covariates are included in the calculation; thus, multipliers c_j (values of 0 or 1) are used to select the covariates for this purpose, which allows adjustment of the model for potential confounders. For a given age and sex, the weighted average relative risk in the controls corresponds to the population average absolute risk given by SEER; therefore, for an individual with specific covariate values x_j giving a relative risk estimated by ${\text{e}}^{{\mathrm{\Sigma }}_j{c}_j{\beta }_j{x}_j},$ the estimated absolute risk is obtained by multiplying his or her SEER absolute risk (SEER in the equations below) by

$$\frac{{\text{e}}^{{\sum }_j{c}_j{\beta }_j{x}_j}}{\text{Average}\text{relative}\text{risk}}=\frac{{\text{e}}^{{\sum }_j{c}_j{\beta }_j{x}_j}}{{\displaystyle (1/n){\sum }_i{w}_i{\text{e}}^{{\sum }_j{c}_j{\beta }_j{x}_{ij}}}}=\frac{1}{{\displaystyle (1/n){\sum }_i{w}_i{\text{e}}^{{\sum }_j{c}_j{\beta }_j(x_{ij}-x_j)}}}=\frac{1}{D},$$

where the denominator $D={\displaystyle (1/n){\sum }_i{w}_i{\text{e}}^{{\sum }_j{c}_j{\beta }_j(x_{ij}-x_j)},}$ that is, the individual's absolute risk = SEER/D. Because the SEER data and the parameter estimates in D are independent, the delta method (19) gives var(absolute risk) = var(SEER/D) = [SEER² var(D) + D² var(SEER)]/D⁴. The DevCan program for the SEER probabilities provides their 95% confidence limits, which then yield var(SEER). All that remains is to calculate var(D). This is also done using the delta method by taking $f_i={\displaystyle {\text{e}}^{{\sum }_{\ell }{c}_{\ell }{\beta }_{\ell }(x_{i\ell }-x_{\ell })}:}$

$$\begin{array}{cc}\text{var}(D)\hfill & =\frac{1}{n^2}\left\{\sum _j{\left[\sum _i{w}_i{c}_j(x_{ij}-x_j)f_i\right]}^2\text{var}({\beta }_j)\right.\left.+2\sum _{k<j}\left[\sum _i{w}_i{c}_j(x_{ij}-x_j)f_i\right]\left[\sum _i{w}_i{c}_k(x_{ik}-x_k)f_i\right]\text{cov}({\beta }_j,{\beta }_k)\right\}.\hfill \end{array}$$

Confidence limits for the absolute risk are then obtained from its calculated variance.

RESULTS

Distributions of the various factors among cases and controls are given in Table 1. Cases were on average 1–2 years older than controls and included slightly larger percentages of other races than of whites. Body mass indices (weight in kilograms divided by the square of height in meters) were similar for cases and controls, although slightly more cases were obese (body mass index >30) than controls, especially among men. History of pancreatic cancer among first-degree relatives, non-O blood group, Jewish ancestry, and H. pylori seropositivity were more prevalent among cases. Cases also smoked more and longer than controls, were more likely to be current smokers, and had quit for less time. History of diabetes mellitus, pancreatitis, and regular use of proton-pump inhibitors were all more frequent among cases than among controls and occurred at older ages. Ever regular use or current use of histamine receptor-antagonist medications and of antacids did not show substantial case-control differences.

Table 1.

Characteristics of Pancreatic Cancer Case Patients and Normal Population Control Subjects, Connecticut, 2005–2009

Characteristic	Men						Women
	Cases			Controls			Cases			Controls
	No.	%	Mean (SD)	No.	%	Mean (SD)	No.	%	Mean (SD)	No.	%	Mean (SD)
Total no. of subjects	207			390			155			300
Age at interview, years	45.7–85.7a		68.1 (9.2)	36.6–84.3a		67.0 (10.2)	37.5–85.1a		68.9 (10.6)	36.0–85.1a		66.8 (10.7)
Race
White	190	91.8		372	95.4		146	94.2		284	94.7
Other	17	8.2		18	4.6		9	5.8		16	5.3
Jewish ancestry
No	185	89.4		362	92.8		137	88.4		279	93.0
Yes	22	10.6		28	7.2		18	11.6		21	7.0
ABO blood group
O	84	40.6		176	45.1		61	39.4		139	46.3
Other	123	59.4		214	54.9		94	60.6		161	53.7
History of pancreatic cancer in first-degree relatives
No	194	93.7		371	95.1		144	92.9		289	96.3
Yes	13	6.3		19	4.9		11	7.1		11	3.7
BMIb			27.5 (4.8)			27.7 (4.6)			26.4 (5.4)			26.8 (6.4)
No. of subjects with BMI >30	53	25.6		82	21.0		37	23.9		70	23.3
H. pylori seropositivity
No	164	79.2		316	81.0		121	78.1		254	84.7
Yes	43	20.8		74	19.0		34	21.9		46	15.3
H. pylori CagA seropositivity
No	178	86.0		327	83.8		130	83.9		255	85.0
Yes	29	14.0		63	16.2		25	16.1		45	15.0
History of diabetes mellitus
No	139	67.1		317	81.3		118	76.1		270	90.0
Yes	68	32.9		73	18.7		37	23.9		30	10.0
Age at diabetes diagnosis, years			60.0 (11.7)			55.1 (13.4)			61.5 (13.2)			57.4 (13.0)
Time before pancreas cancer diagnosis, years			9.7 (8.4)			12.8 (11.9)			9.2 (10.9)			13.2 (12.0)
No. diagnosed within previous 2 yearsc	7	10.3		6	8.2		4	10.8		1	3.3
History of pancreatitis
No	193	93.2		378	96.9		138	89.0		297	99.0
Yes	14	6.7		12	3.1		17	11.0		3	1.0
Age at pancreatitis diagnosis, years			60.1 (11.1)			53.8 (15.6)			63.8 (10.8)			55.7 (15.0)
Time before pancreas cancer diagnosis, years			7.8 (10.2)			9.2 (10.2)			4.1 (8.3)			9.0 (8.0)
No. diagnosed within previous 2 yearsc	8	57.1		5	41.7		13	76.5		1	33.3
Regular cigarette smoking
Never	55	26.6		140	35.9		57	36.8		146	48.7
Current	47	22.7		43	11.0		36	23.2		38	12.7
Past	105	50.7		207	53.1		62	40.0		116	38.7
Smoking duration,d years			31.6 (15.7)			26.9 (15.1)			31.4 (16.3)			25.3 (15.5)
Cigarettes smoked per dayd			29.3 (17.1)			25.5 (16.4)			22.2 (13.8)			19.3 (11.7)
Pack-years of smokingd			48.2 (37.1)			36.9 (34.5)			37.9 (34.2)			26.1 (22.3)
Time since quitting,d years			17.3 (17.3)			21.8 (15.3)			15.8 (17.2)			19.4 (15.4)
No. who quit within previous 2 yearsc	50	32.9		44	17.6		37	37.8		38	24.7
Regular use of PPIs
Never	154	74.4		329	84.4		106	68.4		253	84.3
Ever	53	25.6		61	15.6		49	31.6		47	15.7
Currente	49	23.7		51	14.1		42	27.1		35	11.7
Duration of PPI use,e years			4.1 (4.2)			5.9 (6.3)			3.9 (5.4)			4.5 (4.3)
Age at starting,e years			63.9 (9.0)			59.6 (12.3)			64.2 (11.6)			60.9 (11.7)
Time from starting to interview,e years			5.7 (4.7)			9.1 (9.5)			5.7 (6.5)			7.4 (7.1)
No. who started within previous 2 yearsc	8	15.1		2	3.3		8	16.3		6	12.8
Regular use of H2RAs
Never	189	91.3		370	94.9		147	94.8		278	92.7
Ever	18	8.7		20	5.1		8	5.2		22	7.3
Currente	12	5.8		9	2.3		5	3.2		8	2.7
Duration of H2RA use,e years			5.3 (6.3)			7.3 (9.3)			6.0 (8.9)			4.6 (4.9)
Age at starting,e years			56.2 (13.9)			52.3 (12.7)			50.8 (18.0)			51.1 (13.8)
Time from starting to interview,e years			10.4 (12.6)			14.4 (11.0)			11.9 (10.8)			16.0 (12.6)
Regular use of antacids
Never	178	86.0		359	92.1		135	87.1		261	87.0
Ever	29	14.0		31	7.9		20	12.9		39	13.0
Currente	18	8.7		17	4.4		11	7.1		22	7.3
Duration of antacid use,e years			11.7 (13.3)			16.3 (15.9)			7.6 (9.8)			7.6 (8.1)
Age at starting,e years			51.2 (15.4)			46.0 (15.7)			54.3 (9.8)			52.1 (13.5)
Time from starting to interview,e years			16.5 (13.5)			21.1 (13.1)			14.2 (12.9)			12.8 (7.7)

Abbreviations: BMI, body mass index; CagA, cytotoxin-associated gene A; H. pylori, Helicobacter pylori; H₂RAs, histamine receptor-antagonist heartburn medications; PPIs, proton pump-inhibitor medications; SD, standard deviation.

^a Values are expressed as a range.

^b Usual adult weight in kilograms divided by the square of height in meters.

^c Within 2 years of interview.

^d Among ever-smokers.

^e Among ever-users.

Table 2 shows parameter estimates and odds ratios for the factors used in predicting absolute 5-year risks, including the 4 time variables in the exponential time-decay model β₁I exp(−β₂t), as described above. In our data, not all of the time parameters had statistically significant parameter values at a 2-sided P value of 0.05; however, they have been retained in the model because of evidence from the literature that the magnitudes of risk decline over time. Figure 1 gives graphical presentations of the decline in the odds ratio of pancreatic cancer by how many years in the past diabetes was diagnosed, cigarette smoking ceased, proton-pump inhibitor use started, and pancreatitis occurred. In each of the figure panels, we also show the odds ratio and its 95% confidence limits according to strata of 0–2 years in the past and additional past intervals when the factor occurred. The continuous odds ratio in each panel is seen to lie within the category 95% confidence limits in all cases and generally does a good job of representing the category odds ratios over past time.

Table 2.

Parameter Estimates and Relative Odds of Pancreatic Cancer for Recent Factor Occurrence and for Occurrence 5 Years in the Past as Determined by Using a Log Odds Ratio Exponential Time-Decay Model, Connecticut, 2005–2009

Factor	Parameter Estimate	Standard Error	P Value	Time of Occurrence	ORa,b	95% CIb
Jewish ancestry	0.591	0.275	0.032	Ever	1.81	1.05, 3.10
Non-O ABO blood group	0.271	0.160	0.090	Ever	1.31	0.96, 1.79
Diagnosis of diabetes mellitus	1.57	0.400	10−4.1	Current	4.81	2.19, 10.6
Time since diabetes diagnosis, year−1	0.062	0.038	0.10	5 years in the past	3.16	2.07, 4.82
Diagnosis of pancreatitis	2.97	0.942	0.0016	Current	19.5	3.08, 124.
Time since pancreatitis diagnosis, year−1	0.32	0.26	0.22	5 years in the past	1.84	0.50, 6.95
Current cigarette smoking	1.181	0.217	10−7.3	Current	3.26	2.13, 4.98
Time since quitting smoking, year−1	0.092	0.047	0.050	5 years in the past	2.11	1.36, 3.28
Current use of PPIs	1.83	0.667	0.0061	Current	6.21	1.68, 22.9
Time since starting use of PPIs, year−1	0.18	0.11	0.10	5 years in the past	2.08	1.31, 3.32

Abbreviations: CI, confidence interval; OR, odds ratio; PPIs, proton-pump inhibitor heartburn medications.

^a Odds ratio with respect to nonexposure to the factor. Adjusted for age, sex, white versus other race, history of pancreatic cancer in first-degree relatives, seropositivity for Helicobacter pylori and cytotoxin-associated gene A, as well as for all of the factors in the table.

^b For the time variables, the parameter estimate is β₂ in the model term β₁I exp(−β₂t); the odds ratio and confidence interval are for the exposure factor occurring 5 years before diagnosis of pancreatic cancer. β₁ is given by the parameter estimate in the immediately preceding line.

Figure 1.

Estimated odds ratio for pancreatic cancer in a case-control study, Connecticut, 2005–2009. The odds ratio is calculated as a continuous value according to the model formula exp{β₁I exp(−β₂t)}, where I denotes personal history of the particular factor (0 or 1) and t its time-in-the-past variable (in years). Shown in the figure is the continuous odds ratio, along with the odds ratio of the factor grouped according to categories of years before interview when the factor occurred. A) Diagnosis of diabetes mellitus; B) current cigarette smoking or smoking cessation; C) start of proton pump-inhibitor use; and D) diagnosis of pancreatitis. The thick continuous dashed lines indicate the estimated model odds ratios; dotted continuous lines, 95% confidence limits for the model odds ratios; thick black horizontal step lines, estimated odds ratios for the categories of years before interview; and shaded boxes, 95% confidence limits for the category odds ratios. Category widths were chosen in order to obtain approximately equal numbers of control subjects in each group.

Table 3 gives estimated absolute risks of developing pancreatic cancer over 5 years by categories of age and sex according to various combinations of the risk factors listed in Table 2. These risks are estimated directly as SEER/D, as outlined above. At all ages, lifelong nonsmokers of non-Jewish ancestry and with O blood type who had never been diagnosed with diabetes or pancreatitis or regularly used proton-pump inhibitors had 5-year risks of less than 0.15%. Such unexposed individuals are infrequent, comprising 3.3% of cases and 13% of controls. Various combinations of the risk factors led to estimated risks that reached 5%–10% in the older age groups. These combined-factor risk scenarios were found among our study subjects and are not extrapolations. For example, smokers who had had pancreatitis comprised 1.3% of controls and 4.4% of cases, and proton-pump inhibitor users with diabetes comprised 2.6% of controls and 8.8% of cases. In total, 0.87% of controls had estimated 5-year absolute risks of 5% or higher.

Table 3.

Estimated Absolute Risk of Developing Pancreatic Cancer Within the Next 5 Years, 13 Surveillance Epidemiology and End Results Registries, 2008–2010, and Connecticut Study Data, 2005–2009

Starting Age, years	Unexposed		Diabetes, PPI Use, Nonsmoker, Blood Group O, Non-Jewish Ancestry		Diabetes, PPI Use, Nonsmoker, Blood Group Non-O, Jewish Ancestry		Diabetes, PPI Use, Current Smoker, Blood Group Non-O, Non-Jewish Ancestry		Current Smoker, Recent Pancreatitis, Blood Group Non-O, Non-Jewish Ancestry
	Absolute Risk, %	95% CI	Absolute Risk, %	95% CI	Absolute Risk, %	95% CI	Absolute Risk, %	95% CI	Absolute Risk, %	95% CI
	Men
50	0.021	0.013, 0.029	0.62	0.38, 0.87	0.88	0.48, 1.28	2.03	1.24, 2.82	1.43	0.92, 1.94
55	0.036	0.022, 0.050	1.08	0.66, 1.50	1.53	0.84, 2.23	3.52	2.15, 4.89	2.48	1.61, 3.36
60	0.058	0.035, 0.080	1.72	1.05, 2.39	2.44	1.35, 3.54	5.61	3.43, 7.79	3.95	2.56, 5.35
65	0.082	0.050, 0.114	2.46	1.51, 3.41	3.49	1.92, 5.06	8.01	4.90, 11.12	5.64	3.65, 7.63
70	0.102	0.062, 0.141	3.04	1.86, 4.22	4.31	2.37, 6.25	9.90	6.06, 13.75	6.97	4.51, 9.43
75	0.124	0.076, 0.172	3.70	2.26, 5.14	5.25	2.89, 7.61	12.05	7.37, 16.73	8.48	5.49, 11.48
80	0.147	0.090, 0.204	4.37	2.67, 6.08	6.21	3.41, 9.00	14.25	8.71, 19.79	10.03	6.49, 13.58
	Women
50	0.015	0.010, 0.021	0.46	0.29, 0.63	0.65	0.36, 0.94	1.50	0.96, 2.04	1.05	0.71, 1.40
55	0.026	0.017, 0.036	0.79	0.51, 1.07	1.12	0.63, 1.61	2.57	1.65, 3.49	1.81	1.23, 2.39
60	0.044	0.028, 0.059	1.31	0.84, 1.77	1.85	1.04, 2.67	4.26	2.75, 5.77	3.00	2.04, 3.95
65	0.064	0.041, 0.087	1.92	1.24, 2.60	2.72	1.53, 3.92	6.25	4.03, 8.47	4.40	3.00, 5.80
70	0.094	0.060, 0.127	2.79	1.80, 3.78	3.96	2.23, 5.70	9.10	5.87, 12.32	6.41	4.37, 8.44
75	0.122	0.079, 0.165	3.65	2.35, 4.94	5.17	2.91, 7.44	11.87	7.67, 16.08	8.36	5.71, 11.01
80	0.141	0.091, 0.190	4.20	2.71, 5.68	5.95	3.35, 8.56	13.67	8.83, 18.52	9.63	6.58, 12.68

Abbreviations: CI, confidence interval; PPI, proton pump-inhibitor heartburn medication.

DISCUSSION

The present analysis has a number of strengths. We used population risk data collected from 13 US SEER registries spanning the years 2008–2010, including 6,289 men and 5,865 women diagnosed with pancreatic cancer. Our case-control study, from which we calculated estimates of relative risk, involved a large and representative sample of Connecticut population subjects. We obtained comprehensive questionnaire information on a variety of both known and suspected risk factors and sought detailed information about the timing of past events.

Some potential limitations of our analysis should also be considered. Our study response rates were acceptable for a case-control study, but as in any such subject sample, they still could have led to some bias in parameter estimates compared with what would have been observed in the full underlying population. Given the poor survival rate in pancreatic cancer, our case sample was likely skewed toward cases with earlier stage disease rather than being representative of all newly diagnosed individuals. Our total number of cases (362), although substantial, is too small to subdivide the risk-factor analysis according to stage at diagnosis. Thus, our observed magnitudes of associations might apply to a certain cross-section of people with pancreatic cancer as opposed to only individuals with early-stage disease, which would have been desirable for predicting the occurrence of new diagnoses. Our risk factor questionnaire did not distinguish between reported episodes of acute and chronic pancreatitis, forcing us to consider any history of pancreatitis as a single exposure even though acute and chronic pancreatitis might have some inherently different risks. We did not ascertain subjects’ recent body weights and therefore could not incorporate unintentional weight loss in our model. Also, the confidence limits of our estimated absolute risks tended to be appreciable, although not so large as to change any conclusions about whether the risks would be sufficient to motivate diagnostic workup among individuals with the various factors. Finally, we assumed that the odds ratios of the various factors in our model did not vary by age; that is, we did not include age interactions in the model. Our sample size was not large enough to determine whether the magnitude of the odds ratio varied by age group.

In general, certain combinations of the factors examined in the present study could be used to identify small segments of the older adult population with substantially higher risks of pancreatic cancer. These risks are highest at the time that a new factor occurs or within approximately 2 years of occurrence, and they decline appreciably over additional time. The magnitudes of risk, which are in the 5%–10% range, are similar to those seen in clinic populations of individuals with large familial risks (20). Although of very appreciable magnitudes, these large risks account for small fractions of all pancreatic cancer diagnoses, and thus the factors in the present model might provide for only a beginning approach to studying risks of new-onset disease. Nevertheless, the factors in our model appear to predict risk better than newly diagnosed diabetes on its own or in conjunction with various additional factors (21–23).

The parameter values of our model can be validated by comparison with magnitudes of risks seen in studies elsewhere, including in large meta-analyses. Table 4 presents comparisons of the odds ratios in our analysis with values from the literature and shows a general similarity between them. Ashkenazi Jews have a risk of pancreatic cancer that is approximately 50% higher than that in the general population even after adjustment for known risk factors (24). Our odds ratio of 1.81 is approximately 25% larger than the 50% elevation, and thus the actual risks for the middle scenario in Table 3 might be approximately 80% of the table values. For pancreatitis, our estimated odds ratios were smaller than those seen in a large meta-analysis (25); however, our risks for current smoking were somewhat larger than what has been calculated in 2 meta-analyses (6, 26), and these opposing associations largely cancel each other in the last scenario in Table 3. Use of proton pump-inhibitors more than 2 years in the past was examined in 1 record-linkage study from the United Kingdom (27), but use within 2 years of pancreatic cancer diagnosis, which could be reflective of cancer symptoms, was not examined. In any event, if we had used the odds ratios from the literature that are shown in Table 4 to estimate absolute risks, the ranges of risk in Table 3 would remain between 5%–10% for older individuals.

Table 4.

Comparison of Parameter Estimates in Present Model With Literature Values

Present Model		Literature
Odds Ratioa	Duration in Pastb	Odds Ratio	Duration in Pastb	First Author, Year (Reference No.)
Jewish Ancestry
1.81		1.43		Eldridge, 2011 (24)
ABO Non-O Blood Group
1.31		1.41c		Risch, 2013 (4)
		1.31d		Risch, 2013 (4)
Diabetes Mellitus
4.81	Recente	3.87	0–1 year	Chari, 2008 (28)
4.32	0–2 years	4.2f	0–2 years	Mizuno, 2013 (29)
		5.38	<1 year	Ben, 2011 (30)
		1.95	1–4 years	Ben, 2011 (30)
		1.49	5–9 years	Ben, 2011 (30)
		1.6	6–10 years	Li, 2011 (31)
		1.47	≥10 years	Ben, 2011 (30)
		1.4	≥11 years	Li, 2011 (31)
Pancreatitis
19.5	Recent
9.76	0–2 years	13.6	0–2 years	Duell, 2012 (25)
2.10	>2 years	2.71	>2 years	Duell, 2012 (25)
Cigarette Smoking
3.26	Current	2.20	Current	Bosetti, 2012 (6)
3.26	Quit 0–2 years	1.74	Current	Iodice, 2008 (26)
		1.64	Quit 1–10 years	Bosetti, 2012 (6)
		1.42	Quit 10–15 years	Bosetti, 2012 (6)
		1.15	Quit ≥10 years	Iodice, 2008 (26)
Use of PPIs
6.21	Recent
5.15	0–2 years
1.87	>2 years	1.02	>2 years	Bradley, 2012 (27)

Abbreviation: PPIs, proton-pump inhibitors.

^a Adjusted as in Table 2.

^b Blank entries are not time-dependent.

^c In Western populations.

^d In combined Western and Asian populations.

^e Odds ratio with respect to nonexposure to the factor for recent occurrence (i.e., t = 0 in past).

^f Three-fold increased risk compared with long-duration adult-onset diabetes mellitus, and assuming that long-duration diabetes conveys 1.4-fold risk.

In the present analysis, by combining known risk factors for pancreatic cancer and detectable prediagnostic symptomatology, we have begun to identify small segments of the population that have risks sufficiently high enough to make screening potentially useful. Our observation of an association of higher risk with recent usage of proton-pump inhibitors is a novel finding; only 1 study of pancreatic cancer to date has obtained detailed information on heartburn or acid reflux medications, but it excluded analysis of use within 2 years of diagnosis. Ultimately, it seems entirely possible that further characterization of early signs and symptoms and increasingly accurate estimates of their risk associations in combination with known risk factors for pancreatic cancer will yield models able to delineate absolute risks great enough to warrant screening for the disease.