Development of a Clinical Prediction Model for Diabetes in Chronic Pancreatitis: The PREDICT3c Study

Chronic pancreatitis (CP) is a highly morbid condition often characterized by pain and maldigestion of nutrients. One of the most common complications of CP is development of CP-associated diabetes mellitus (CP-DM). In cross-sectional analyses of adults with CP, the prevalence of CP-DM is consistently in the range of 30–40% (1,2). In longitudinal studies, the cumulative prevalence of diabetes has been reported to be 15–26% at 10 years, 46–83% over 20 years, and as high as 90% at 50 years after onset of CP (3–7). A meta-analysis of 15 studies (including 8,970 people with CP) documented incidence rates of new-onset diabetes of 15% within 3 years and 33% after 5 years (8). Among nearly 32,000 adults with newly diagnosed diabetes, 97% were found to have type 2 diabetes, 1.1% to have type 1 diabetes, and 1.6% to have diabetes after pancreatic disease; of the latter, 16% had CP-DM (9).

CP-DM is characterized by earlier progression to insulin treatment compared with type 2 diabetes (9). In addition to experiencing risk of the usual microvascular complications of diabetes (retinopathy, nephropathy, neuropathy) (10), patients with CP-DM are at increased risk for severe hypoglycemia, heightened all-cause mortality (11,12), and other complications. While both type 2 diabetes and CP are associated with an increased risk of pancreatic cancer, individuals with both CP and diabetes are at the highest risk (13–15). Overall cancer mortality is increased in CP-DM relative to type 2 diabetes, particularly in women (16).

Given this additional morbidity directly and indirectly related to diabetes in those with CP, methods to predict and prevent diabetes in CP are urgently needed. The first step in this process is to identify independent risk factors associated with diabetes in CP. As recently reviewed (17), several studies investigating such factors have identified both traditional risk factors for type 2 diabetes (e.g., older age, obesity, family history of diabetes) and pancreatic disease–specific factors (e.g., exocrine dysfunction, pancreatic calcification). As several risk factors are correlated with each other (e.g., smoking and alcohol use), multivariable analysis using patient-level data are needed to minimize confounding, which had been performed in some but not all studies. Herein, we present comprehensive analyses of risk factors for diabetes in CP in a large, well-phenotyped cohort of patients with CP from the Prospective Evaluation of Chronic Pancreatitis for Epidemiologic and Translational Studies (PROCEED) cohort (18). We give special attention to the timing of diabetes in relation to a CP diagnosis to interrogate the factors associated with pancreatogenic diabetes. With the identified risk factors, we aimed to develop a prediction model for diabetes in patients with CP and evaluate the model by cross validation.

The PROCEED study is an ongoing, multicenter, longitudinal cohort study of CP in the U.S. (ClinicalTrials.gov identifier: NCT03099850) sponsored by the National Institutes of Health (18). The study was approved by the institutional review boards of each of the participating institutions and the coordinating and data management center of the Consortium for the Study of Chronic Pancreatitis, Diabetes, and Pancreatic Cancer (CPDPC). Informed consent was provided by participants before any study procedures.

In this study, we conducted a cross-sectional analysis of data available at baseline. Patients included in this analysis had definite CP, were aged 18–75 years, and were enrolled in PROCEED from June 2017 to August 2021 (n = 669). CP was defined by at least one of the following criteria: 1) pancreatic parenchymal and/or ductal calcifications and/or Cambridge stage 3–4 findings on cross-sectional imaging or 2) histologic evidence of CP (19). Cross-sectional imaging was reviewed by a CPDPC site radiologist to confirm the diagnosis, using a standardized approach (20). Patients with CP were identified from gastroenterology clinics.

As previously reported, detailed demographic and clinical data were prospectively collected from participants using standardized questionnaires and from physicians using structured case report forms (18). Of 669 participants with definite CP, complete assessment of diabetes status was available for 645. For those with known diabetes before enrollment, the method (e.g., hemoglobin A_1c [HbA_1c], blood glucose levels) and date of first diagnosis, if known, was recorded, along with current use of antidiabetic medications. Participants with no known history of diabetes underwent a per-protocol assessment for diabetes, including blood glucose (fasting or random) and HbA_1c levels, using American Diabetes Association criteria, i.e., abnormal values on two of the following tests or two abnormal values of the same test: fasting blood glucose ≥126 mg/dL, HbA_1c ≥6.5%, random blood glucose ≥200 mg/dL. From these assessments, we determined that 276 participants had prevalent diabetes. Of these participants, 258 (93.5%) were enrolled with a prior physician or had a laboratory diagnosis of diabetes, and 18 (6.5%) were identified by abnormal glucose or HbA_1c parameters per protocol.

On the basis of the timing of diabetes diagnosis in relation to the time of diagnosis of pancreatitis (defined as the first diagnosis of acute pancreatitis or CP, whichever occurred earliest), we stratified the participants with pancreatitis and diabetes into three groups: 1) the prepancreatitis DM group included participants whose diabetes was diagnosed at least 1 year before the first diagnosis of pancreatitis, 2) the concurrent DM group included participants whose diabetes was diagnosed within 1 year of pancreatitis, and 3) the postpancreatitis DM group included participants whose diabetes was diagnosed at least 1 year after their pancreatitis diagnosis. We also performed an analysis comparing the prepancreatitis DM group with a combination of the concurrent DM and postpancreatitis DM groups.

We abstracted information on patient and disease-related factors that may be associated with the odds of having prevalent diabetes on the basis of published literature and other potential factors from the PROCEED database. Tobacco use (cigarette smoking, tobacco chewing, cigar smoking) was classified as never, past, or current. Intensity of cigarette smoking was reported as packs per day. Alcohol use was classified as never, past, or current. With regard to intensity of alcohol consumption, we categorized participants as abstainers (no alcohol use or <20 drinks in a lifetime), light drinkers (≤3 drinks/week), moderate drinkers (4–7 drinks/week for females, 4–14 drinks/week for males), heavy drinkers (8–27 drinks/week for females, 15–34 drinks/week for males), and very heavy drinkers (≥28 drinks/week for females, ≥35 drinks/week for males) using drinking information during the heaviest drinking period and following definitions applied in a prior cross-sectional study of CP (21). Pancreatitis-related factors included history of acute pancreatitis (ever, recurrent acute pancreatitis, complications of acute pancreatitis such as organ failure, necrosis, fluid collection), age at first diagnosis of pancreatitis and etiology based on physician assessment (alcoholic, nonalcoholic), and history of endoscopic therapy or pancreatic surgery (resection, drainage procedure, other). Presence of exocrine pancreatic dysfunction (EPD) before enrollment and the method of diagnosis, if known (fecal elastase or chymotrypsin, quantitative fecal fat, clinical signs of steatorrhea) was noted. Participants with no known EPD underwent a per-protocol assessment with fecal elastase; a value of <100 μg/g stool was considered as positive for EPD. The status of EPD for participants with no known history and who did not undergo a per-protocol assessment was designated as not tested. Information on the details of cross-sectional imaging studies (i.e., computed tomography, MRI) were retrieved from the case report forms completed by CPDPC site radiologists (e.g., pancreatic calcifications, pancreatic atrophy). Presence or absence of common bile duct or pancreatic duct stricture was determined by findings on imaging or during endoscopy or surgery.

A univariable analysis with diabetes (yes, no) status as the dependent variable was performed on all described variables to select candidates for multivariable analysis at a significance level of P ≤ 0.2. P values for categorical variables were determined using χ² or Fisher exact tests, as appropriate. P values for numeric variables were determined using the nonparametric Wilcoxon rank sum test. Logistic regression was applied on the selected variables to determine their independent effects on the presence of diabetes in multivariable regression models.

We assessed the predictive accuracy of each model for identifying diabetes with two cross-validation methods. First, 2,000 bootstrap samples were created from the original data set and used as training data sets, and the corresponding out-of-bootstrap samples were used as internal validation data sets. The predictive accuracy measure as area under the receiver operating characteristic curve (AUROC) was obtained by averaging 200 curves produced independently using bootstrap internal cross validation. The AUROC measures how well the model can correctly distinguish the patients with CP-DM from patients with CP but without diabetes. A higher AUROC value indicates better predictive performance of the model and, hence, stronger overall association between the risk factors and diabetes outcomes. Second, we performed a by-institution cross validation. We used eight institutions of the PROCEED study for training and one for testing. The predictive accuracy measure was obtained using a patient population–weighted average over the nine-way AUROC split. Because the nine institutions do not have the same patient population, this by-institution cross validation has better generalizability than the bootstrap internal cross validation.

All these analyses were also performed on three subsets of participants stratified on the basis of the timing of diabetes diagnosis in relation to the time of pancreatitis (prepancreatitis DM, concurrent DM, and postpancreatitis DM). Univariable and multivariable analysis for these three mutually exclusive subgroups were performed using the same reference group of 369 patients with CP but without diabetes. We combined underweight and normal weight (according to BMI) patients in the prepancreatitis DM subgroup to avoid having zero cell frequency in the underweight with diabetes category. Furthermore, pancreatitis duration at onset of diabetes was excluded from subgroup multivariable analyses because it is the primary characteristic used to stratify these subgroups. Multivariable analysis was also similarly performed on two subsets of participants (prepancreatitis DM and a combination of concurrent DM and postpancreatitis DM).

No participants with CP were excluded from the analysis because of missing data as long as their diabetes status was known (yes, no). Missing data accounted for <2% of the observations in the worst case. Therefore, single mean imputation was applied to missing numeric variable data, and single-mode imputation was applied to missing categorical variable data. Unless stated otherwise, P < 0.05 was used as the level of statistical significance.

We followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guidelines (22). In accordance with TRIPOD, Supplementary Note 1 describes the prediction model in detail, Supplementary Table 1 displays the clinical features per diabetes type, Supplementary Table 2 presents the raw coefficient values for the multivariate regression models, and Supplementary Figs. 1–5 display the ROC curves for the models included in this study.

Samples and data are available to the wider scientific community in accordance with the National Institutes of Health policy on data sharing, as well as the National Institute of Diabetes and Digestive and Kidney Diseases policy for data sharing in multicenter clinical studies.

We studied 645 patients with a confirmed CP diagnosis and a complete assessment of diabetes, of whom 276 had diabetes and 369 did not at baseline. We initially conducted univariable analyses examining the association between each risk factor and prevalent diabetes (Table 1). Older age, being overweight or obese, male sex, non-White race, history of ≥20 pack-years of cigarette smoking, common bile duct stricture, history of pancreatic surgery, family history of diabetes, EPD, pancreatic calcifications, and pancreatic atrophy were associated with prevalent diabetes (P < 0.05). Borderline significant associations with diabetes were observed for duration of pancreatitis, Hispanic ethnicity, tobacco use, nonalcoholic etiology of pancreatitis, and history of acute pancreatitis complications. The remaining variables were not associated with diabetes.

The variables with significant or nearly significant associations (P ≤ 0.2) with diabetes were next analyzed jointly in multivariable regression (Table 2). Factors that remained significantly associated with diabetes were older age, overweight/obese status, male sex, non-White race, current tobacco use, history of acute pancreatitis complications, nonalcoholic etiology, presence of EPD, not being tested for EPD, pancreatic calcifications, and pancreatic atrophy.

Next, we repeated the multivariable regression for three different diabetes definitions (prepancreatitis DM [n = 68], concurrent DM [n = 62], and postpancreatitis DM [n = 106]) according to when the diabetes diagnosis was made in relation to the first diagnosis of pancreatitis. As shown in Table 3, overweight/obese status, male sex, and nonalcoholic etiology were associated with an increased prevalence of diabetes in all three groups. Hispanic ethnicity, common bile duct stricture, and EPD were associated with prepancreatitis DM and concurrent DM, and acute pancreatitis complications and atrophy were associated with concurrent DM and postpancreatitis DM. Age, other race (non-Black vs. White), and family history of diabetes were associated with only prepancreatitis DM. Smoking history of ≥20 pack-years and EPD not tested were associated with diabetes only in the concurrent DM group. Black race, pancreatic surgery, and pancreatic calcifications were associated with only postpancreatitis DM.

The performance of these multivariable models was assessed by AUROC analysis. In the main analysis of all participants, the AUROC was 0.745 (95% CI 0.693, 0.796). In the stratified analysis of prepancreatitis DM, the AUROC was 0.809 (95% CI 0.715, 0.890). For concurrent DM, the AUROC was 0.736 (95% CI 0.614, 0.830). For postpancreatitis DM, the AUROC was 0.705 (95% CI 0.624, 0.777). These AUROC values indicate good ability of the models to discriminate CP with versus without diabetes. By-institution cross validation yielded similar performance as the bootstrap internal cross validation, with AUROC values of 0.755, 0.798, 0.751, and 0.717 for the main analysis, prepancreatitis DM, concurrent DM, and postpancreatitis DM, respectively.

Multivariable regression for two diabetes subgroups (prepancreatitis DM and concurrent DM combined with postpancreatitis DM) are presented in Supplementary Table 3. The results were similar to those seen in the analysis of the three subgroups with a few exceptions. Compared with the postpancreatitis DM findings shown in Table 3, EPD (both yes and not tested categories) and ≥20 pack-year smoking history in the combined subgroup were significant. On the other hand, surgical therapy was not significant in the combined group (P = 0.058). The remaining risk factors remained significant. The AUROC of this analysis was 0.734 (95% CI 0.674, 0.794).

In a rigorously characterized cohort of patients with CP from the PROCEED study, we found that independent risk factors for diabetes in CP included traditional risk factors for type 2 diabetes (age, BMI category, sex, race, tobacco use) as well as pancreatic disease–specific factors (acute pancreatitis complications, CP etiology, exocrine function, and imaging morphology). We conducted novel subanalyses wherein patients with CP-DM were stratified according to the time of diabetes diagnosis in relation to the first diagnosis of pancreatitis (acute or chronic), under the premise that diabetes developing after pancreatitis is more likely to represent a different form of diabetes than type 2 diabetes. We assumed that our group of patients who developed diabetes before pancreatitis represents type 2 diabetes. While several risk factors, such as older age, male sex, and nonalcoholic etiology, were associated with both prepancreatitis DM and postpancreatitis DM, we found that several other risk factors performed differently in these groups. The traditional risk factors of older age, Hispanic ethnicity, race other than White or Black, and family history of diabetes were significant correlates of prepancreatitis DM but not of postpancreatitis DM. Of note, EPD was also associated with prepancreatitis DM and concurrent DM, which may reflect that type 2 diabetes pathophysiology might have exocrine underpinnings or that diabetes may contribute to EPD (23). However, because many participants did not have prior testing for EPD, this should be investigated further. Furthermore, the risk factors of acute pancreatitis complications, pancreatic surgery, pancreatic calcification, and pancreatic atrophy were significantly associated with postpancreatitis DM but not with prepancreatitis DM. This indicates that factors reflective of damage to the pancreas may increase the risk of developing diabetes in the setting of CP. These results support the notion that pancreatogenic diabetes occurring after pancreatitis may have different pathophysiologic origins than type 2 diabetes. Yet it remains possible that pancreatitis accelerates the pathway to type 2 diabetes by compromising β-cell function.

It is remarkable that several features reflecting pancreatic damage were independently associated with diabetes when analyzed jointly in multivariable analysis. This suggests that the different markers of pancreatic injury may compromise different functions or pathways that contribute to diabetes. While loss of β-cell mass through pancreatic resection is the most straightforward mechanism contributing to diabetes, only 40% of the participants with CP-DM who had undergone surgery had a procedure involving resection, suggesting that more complex mechanisms are involved. Other aspects important to proper islet function, such as blood supply, inflammation, innervation, and autocrine or paracrine signaling, may be differentially affected by prior complicated acute pancreatitis, EPD, pancreatic surgery, and morphologic changes (including pancreatic calcifications and atrophy). Alternatively, the observation that these features are independent risk factors may arise because not all are concurrently present in all patients with CP-DM.

Table 4 summarizes studies that examined risk factors for diabetes occurring in patients with CP. Two earlier studies (6,24) are not listed because their participants were included in a more recent report (3). Across the studies in Table 4, the most frequently observed independent (i.e., significant in multivariable analyses) risk factors were older age of onset of CP, EPD, pancreatic surgery, and pancreatic calcifications. In our study, most of these were significant in the entire cohort (EPD, pancreatic calcification) or emerged as significant in subanalyses (pancreatic surgery). We did not analyze age at onset of pancreatitis because we found that to be highly collinear with duration of pancreatitis and age at enrollment. Older age was strongly associated with diabetes, while duration of pancreatitis exhibited a marginal association in our cohort. The depth of phenotyping in PROCEED allowed us to identify other risk factors, including some not previously described in the literature (race, ethnicity, pancreatic atrophy). Our study also included traditional risk factors for diabetes, which had previously been investigated in depth by only two other studies (1,32). We assessed the performance of our multivariable models with AUROC analysis; only one of the prior studies in Table 4 included predictive performance metrics (3).

The effect of substance use on the risk of diabetes in CP remains unclear. In the current study, those with an alcoholic etiology of pancreatitis had a significantly lower prevalence of diabetes. We cannot determine whether this reflects a protective effect of alcoholic pancreatitis or an increased risk for nonalcoholic etiologies. In contrast to our results, others have observed a higher risk of diabetes in alcoholic CP versus idiopathic or nonalcoholic CP (5,24–27,31) or no difference in diabetes risk for alcoholic CP versus nonalcoholic CP (4,28). In all these studies except one (5), alcoholic etiology was examined in univariable analyses such that the effect of confounding factors, especially by smoking, was not accounted for (Table 4). Consistent with prior studies (7,33), we found that smoking was an independent risk factor for diabetes; however, smoking was not a significant factor in other studies (3,5,30). One study observed that smokers had an earlier onset of diabetes in CP than nonsmokers (33). A physiologic study that found that smoking may accelerate β-cell function deterioration over time in CP (34), providing a potential mechanism for this relationship. A meta-analysis encompassing nearly 6 million individuals found that cigarette smoking was a risk factor for type 2 diabetes, with a dose-response relationship (35).

Strengths of the current study include the large sample size and systematic assessment of extensive phenotypic information, with data recorded using standardized tools and case report forms across multiple study sites distributed across the U.S. Collection of radiologic, endoscopic, and surgical findings in the systematic assessments of PROCEED gave us access to morphologic features that have been infrequently assessed (common bile duct stricture, pancreatic duct stricture, pancreatic atrophy) as potential risk factors for diabetes, especially in combination with lifestyle factors (Table 4). Another unique aspect of PROCEED is the detailed per-protocol assessment of diabetes, a rigorous feature of our methods that reduced the risk of spectrum bias, given that a substantial portion of the diabetes population in the U.S. remains undiagnosed (36). This provided the complete diabetes status for 96% (645 of 669) of the cohort and allowed us to categorize 86% (236 of 276) of those with diabetes according to timing relative to first diagnosis of pancreatitis. Prior pancreatitis cohorts typically did not collect this information in detail. For example, only half of those with diabetes in the North American Pancreatitis Study 2 (NAPS2) had information regarding the chronologic onset of diabetes in relation to pancreatitis (1).

A limitation of our study is that the sample sizes of patients with CP-DM were relatively modest in the subanalyses stratified by diabetes timing in relation to pancreatitis. This may explain some of the unexpected results, such as EPD and common bile duct stricture being associated with prepancreatitis DM and concurrent DM but not with postpancreatitis DM, although it was associated with postpancreatitis DM combined with concurrent DM. As the cohort follow-up continues, we will accumulate incident diabetes cases, allowing more robust statistical power for predictive modeling. Another drawback to the analysis of EPD is the fact that 256 of the 645 participants had not undergone testing for EPD. We included an indicator for this group in the regression models, without imputation, because lack of a fecal elastase test result is a common situation in clinical practice and research.

This analysis represents the first in a line of investigations in the PREDICT3c study, the goal of which is to characterize a set of risk factors that could be used clinically to identify which patients with CP are at the greatest risk of developing diabetes in the future so that preventive measures can be instituted. The AUROC of the currently analyzed clinical features for identifying CP-DM was 0.745, indicating promising predictive value of the factors analyzed thus far. Future work in PREDICT3c will integrate additional assessments, including genomics and surrogates of insulin homeostasis or direct measurements of pancreatic hormones (e.g., insulin, pancreatic polypeptide) (37). As an example, the addition of genetic markers for type 2 diabetes to these clinical features is expected to improve the predictive ability of the model, given that we previously found that a genetic risk score based on such genetic markers was associated with CP-DM (38). Finally, to truly assess whether such models can be used to predict diabetes, analysis of incident cases (rather than prevalent cases as in the current report) will be needed. Participants in PROCEED are being followed longitudinally. We anticipate that over time, a sufficient number of incident cases will be available for us to validate the current models or develop new models to predict diabetes in CP. This would pave the way for diabetes prevention trials in CP.

Acknowledgments. The authors thank all the individuals who volunteered to participate in PROCEED.

Funding. The research reported in this article was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) and National Cancer Institute under award numbers U01DK108327 (P.A.H., D.L.C.), U01DK108288 (S.S.V.), U01DK108323 (E.L.F.), U01DK108326 (W.E.F.), U01DK108328 (L.L., Y.Y.), U01DK108300 (W.G.P.), U01DK108314 (M.O.G.), U01DK108320 (C.E.F.), U01DK126300 (M.D.B.), U01DK108332 (S.K.V.), and U01DK108306 (D.C.W., D.Y.). M.O.G. was supported by the Eris M. Field Chair in Diabetes Research and NIDDK grant P30DK063491.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The funders had no role in the study design; collection, analysis, and interpretation of data; writing of the report; or decision to submit the manuscript. Officials of the NIDDK reviewed the manuscript for scientific content, and the NIDDK approved the final version of the manuscript.

Duality of Interest. M.D.B. has received research support from Viacyte and Dexcom and is a member of the data safety and monitoring board of Insulet and the advisory board of Ariel. C.E.F. and M.O.G. served on an advisory board for Nestlé. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. C.J., P.A.H., L.L., Y.Y., D.L.C., D.Y., and M.O.G. were involved in the conception, design, and conduct of the study and the analysis and interpretation of the results. E.C., M.D.B., W.E.F., E.L.F., C.E.F., W.G.P., S.K.V., S.S.V., J.S., D.C.W., D.L.C., and D.Y. acquired and analyzed data. C.J. and M.O.G. wrote the first draft of the manuscript. All authors edited, reviewed, and approved the final version of the manuscript. M.O.G. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.