A Clinical Prediction Model to Assess Risk for Pancreatic Cancer among Patients with Pre-Diabetes

Abstract

Background:

Early detection of pancreatic ductal adenocarcinoma (PDA) may improve survival. We previously developed a clinical prediction model among patients with new-onset diabetes to help identify PDAs 6 months prior to the clinical diagnosis of the cancer. We developed and internally validated a new model to predict PDA risk among those newly diagnosed with impaired fasting glucose (IFG).

Methods:

We conducted a retrospective cohort study in The Health Improvement Network (THIN) (1995-2013) from the United Kingdom. Eligible study patients had newly diagnosed IFG during follow-up in THIN. The outcome was incident PDA diagnosed within 3 years of IFG diagnosis. Candidate predictors were factors associated with PDA and/or glucose metabolism.

Results:

Among the 138,232 eligible patients with initial IFG diagnosis, 245 (0.2%) were diagnosed with PDA within 3 years. The median time from IFG diagnosis to clinical PDA diagnosis was 326 days (IQR 120 to 588). The final prediction model included age, BMI, proton pump inhibitor use, total cholesterol, LDL, ALT and alkaline phosphatase. The model achieved good discrimination (area under the curve 0.71 [95%CI: 0.67 to 0.75]) and calibration (Hosmer and Lemeshow goodness of fit test p >0.05 in 17 of the 20 imputed data sets) with optimism of 0.0012662 (95%CI: −0.00932 to 0.0108771).

Conclusions:

We developed and internally validated a sequential PDA prediction model based on clinical information routinely available at the initial appearance of IFG. If externally validated, this model could significantly extend our ability to detect PDAs at an earlier stage.

Introduction

Pancreatic ductal adenocarcinoma (PDA) is the fourth most common cause of cancer death, and is expected to rise to the second most common cause by 2030.¹ The 5-year overall survival for pancreatic cancer is only 8% with a substantial survival advantage for localized cancer. However, more than 80% of PDA cases are diagnosed at advanced stages. As it is not cost-effective to screen the general population for PDA, it is imperative to develop low-cost and low-risk tools to identify those patients most at risk for sporadic PDA prior to the development of overt symptoms.

The association between diabetes mellitus (DM) and PDA has been reported in numerous studies. Approximately 50% of newly diagnosed PDA patients have DM, with the vast majority of these DM cases being diagnosed within 2-3 years before PDA diagnosis.² The 3-year risk of PDA among patients with new-onset diabetes is nearly eight times higher than expected.³ Such PDA-associated diabetes is thought to be a paraneoplastic phenomenon likely caused by yet-to-be-determined cancer-induced humoral mediators.^{2, 4} Because of distinct pathogenetic pathways, important differences in clinical features and etiological factors may exist between such paraneoplastic DM and typical type 2 DM.

We recently developed a novel clinical prediction model capable of discriminating these two types of diabetes with excellent accuracy.⁵ However, that model was designed to be applied at the time of clinical diagnosis of DM which may substantially lag the biological onset of the DM. It would be highly desirable to expand our ability to distinguish pancreatic cancer-associated DM from typical type 2 DM at the pre-clinical phase of DM so that we can shift cancer detection to an even earlier time point. Therefore, we conducted a population-based study to develop and internally validate a clinical prediction model for 3-year PDA risk among individuals with pre-diabetes defined as newly detected impaired fasting glucose (IFG).

Methods

Data Source

The current study was conducted using The Health Improvement Network (THIN). THIN database consists of prospectively collected, computerized medical records from a representative sample of general practices throughout the United Kingdom (UK). Under the National Health Services, 98% of UK residents receive all forms of healthcare coordinated through a general practitioner. Therefore, the medical records kept by their general practitioners contain comprehensive clinical data. The THIN patient population as a whole is representative of the age, gender, and geographic distributions of the UK population.⁶ The information collected in the database includes demographics, prescriptions, clinical diagnoses, sub-specialty consult notes, and hospital discharge diagnoses. The Read Clinical Classification is used to classify medical diagnoses.^{7, 8} Each patient visit is associated with a list of coded diagnoses. Participating practices are required to follow predefined protocols for the recording and transferring of clinical data to THIN. Collected data must meet predefined quality standards in order to be included. Previous studies have also confirmed the excellent quality of clinical information in THIN.^{6, 9} Cancer incidence data in THIN, including those for pancreatic cancer, were comparable to those reported in cancer registries in the UK.¹⁰ This study was approved by the University of Pennsylvania Institutional Review Board and the UK’s Scientific Review Committee.

Study Cohort

Study participants were drawn from a cohort of approximately 12 million patients who started follow-up in over 570 THIN-affiliated general practices in the UK from 1995 to 2013. Eligible patients had a newly diagnosed IFG during their follow-up in THIN, aged at least 35 years at the time of the newly diagnosed IFG and had at least 3 years of follow-up in THIN following the diagnosis of IFG unless they developed PDA during the 3 years following initial IFG diagnosis. IFG was defined as having a laboratory test result of fasting glucose level 100 to 125 mg/dL (5.6 to 6.9 mmol/L).¹¹ We defined our cohort to be over 35 years so that the model could be applicable to the vast majority of those with new-onset diabetes/prediabetes who may be at risk for PDA. The exclusion criteria were: patients without acceptable medical records (i.e., patients with incomplete documentation or out of sequence date of birth, registration date, date of death, or date of exit from the database); subjects who were diagnosed with newly diagnosed IFG within the first year after initiation of follow-up in the database in order to avoid prevalent cases; subjects with a diagnosis of PDA prior to the initial diagnosis of IFG; subjects with a diagnosis of PDA >3 years after the diagnosis of IFG.

Dependent Variable

The dependent variable is incident diagnosis of PDA in the eligible study population over the 3-year period following the newly diagnosed IFG. We chose to limit the outcome ascertainment period to 3 years based on the assumption that subclinical PDAs advanced enough to cause IFG through a paraneoplastic mechanism are unlikely to remain undiagnosed after 3 years.

Independent Variables

We evaluated a comprehensive list of PDA risk factors as well as variables related to glucose metabolism (49 candidate variables in total). These predictors included anthropometric variables, lifestyle factors, medical comorbidities, medications, and laboratory studies and were selected based on existing biological and epidemiological evidence. All variables were available in general practice medical record at the time of newly diagnosed IFG. For laboratory studies, last values at the time of or up to 1 year before newly diagnosed IFG were used.

Sample size and power considerations

We identified 245 cases of PDA within 3 years of newly diagnosed IFG in our cohort. With 7 predictors included in the final model, our sample size was substantially larger than the recommended 10 events per predictor for the derivation of a model, as well as the recommended minimum of 100 events for model validation.¹²

Missing data

We excluded variables with ≥ 60% missingness from evaluation as candidate predictors. We performed multiple imputation using multivariate normal regression (MVN) for candidate predictors with < 60% missingness, imputing a total of 20 datasets.^13-16 The MVN method assumes that the data are missing at random.¹⁷

Statistical Analysis

Given that we are predicting the presence of a subclinical cancer, we developed a diagnostic prediction model using logistic regression. All analyses were performed using Stata 13 (Stata Corp, College Station, TX, USA).

Initial candidate predictor selection was based on univariable analysis adjusted for the duration of follow-up from THIN registration date to IFG diagnosis date. All variables associated with a p-value<0.25 in the univariable analyses were subsequently assessed by multivariable logistic regression. For all continuous variables we assessed normality and linearity. Second degree fractional polynomials were used when there is a nonlinear relationship between the continuous predictors and the outcome.^{18, 19} Predictor selection was conducted in each of the 20 imputed datasets using a backward stepwise approach in the multivariable logistic regression model with p-values of <0.001 and >0.05 as the inclusion and exclusion thresholds, respectively. The final multivariable model included those predictors that were selected in ≥50% of the imputation datasets. One advantage of the backward elimination approach over forward selection is that it takes into consideration the correlations among predictors. The overall estimate of model coefficients were derived from all the imputed data sets combined using Rubin’s rule, accounting for uncertainty in the imputed values.^{14, 15} The final multivariable model was tested for collinearity (i.e., variance inflation factor >10). Interactions such as those between BMI and either PPI use or cholesterol levels were tested in the regression model based on clinical or biological plausibility.

Model discrimination was measured by calculating the area under the receiver operating characteristic curve statistic.²⁰ We used the Hosmer–Lemeshow test (based on deciles)²¹ to statistically evaluate the extent of agreement between the predicted and the observed probabilities.

Internal Validation

In order to avoid overfitting and correct for model optimism, we performed an internal validation using a bootstrapping procedure.²² The bootstrapping in the current study was performed using 100 bootstrap resamples of 44,000 individuals each. Predictor selection and model development procedures were repeated in each sample. Optimism in model discrimination was calculated according to Harrell’s algorithm.²³

Results

From 297,773 individuals with new-onset IFG in THIN, we identified 138,232 eligible patients (Figure 1). Among this cohort, 245 individuals (0.2%) were diagnosed with PDA within 3 years of IFG diagnosis. The median follow-up from IFG detection to PDA diagnosis date was 326 days (IQR 120 to 588).

Figure 1:

Participants flow diagram

We evaluated 49 candidate predictors. Seven variables were excluded from the analyses due to missingness of more than 60% (i.e., AST, amylase, ESR, CRP, HbA1C, uric acid, and urinary microalbumin). Among the 42 remaining candidate variables (Supplementary Table 1), 21 had complete data and 21 (mostly lab values) had <60% missing data and were subjected to the multiple imputation procedure. Of the 42 variables that were analyzed in the univariate logistic regression, 20 variables had a p-value<0.25. Notably, the median (interquartile range [IQR]) age at the time of prediabetes diagnosis was significantly older among those subsequently diagnosed with pancreatic cancer than those without pancreatic cancer (70.6 [63.7-77.9] vs. 63.4 [54.6-72.1] years). In addition, the median (IQR) BMI at the time of prediabetes diagnosis was significantly lower among those subsequently diagnosed with pancreatic cancer than those without pancreatic cancer (26.7 [23.9-30.4] vs. 28.4 [25.4-32.2] kg/m²). Monocytes, neutrophil to lymphocyte ratio (NLR), alkaline phosphatase, ALT, LDL and triglycerides were associated with pancreatic cancer risk in a nonlinear fashion, and squared terms were included in the model (Supplemental Table 1). Of note, within-individual trend in glucose levels before IFG diagnosis was not a useful predictor, neither was the absolute value of initial FG within the IFG range (data not shown).

The full multivariable prediction model is presented in Table 1. This model included age, BMI, PPIs, total cholesterol, LDL, ALT and alkaline phosphatase. The area under the curve of the model was 0.71 (95%CI: 0.67 to 0.75) (Figure 2), and the p-value for the Hosmer and Lemeshow goodness of fit test was >0.05 in 17 of the 20 imputed data sets. Internal validation of the model using bootstrapping procedure revealed minimal optimism of 0.0012662 (95%CI: −0.00932 to 0.0108771). Table 2 presents the sensitivity, specificity and positive predictive value for a wide range of risk thresholds used to determine high-risk group targeted for further definitive PDA screening, assuming perfect performance of definitive diagnostic tests (e.g., EUS, MRI) for all existing PDAs. There is an expected tradeoff between sensitivity and positive predictive value. For example, with the screening risk threshold of >0.1%, up to 66% of all PDA cases diagnosed within 3 years of prediabetes onset will be detected. The number needed to screen for this cutoff would be 380 individuals in order to detect one PDA case. When the high-risk group is defined as those with predicted 3-year PDA risk of at least 0.5%, 81 individuals will need to undergo definitive diagnostic testing in order to detect one PDA case within 3 years. In addition, this strategy will detect 16% of all PDAs cases diagnosed within 3 years of prediabetes onset. With a higher risk threshold of > 1%, the number needed to screening decreases to 21, and the strategy will capture 9% of all PDA cases diagnosed within 3 years of prediabetes onset.

Table 1:

Final multivariable prediction model^*

Predictor	β Coefficient	SE	Odds ratio (OR)	95%CI	P-value
Anthropometry
Age	0.045648	0.0074286	1.05	1.03-1.06	<0.001
BMI	−0.0386336	0.0148288	0.96	0.93-0.99	0.012
Medication use
PPIs	0.4307414	0.2344925	1.54	1.14-2.07	0.005
Laboratory studies
Total cholesterol (mmol/L)	−0.2418048	0.1040687	0.79	0.61-1.02	0.068
LDL (mmol/L)	0.1995238	0.3666574	1.22	0.68-2.20	0.507
LDL2	0.0101408	0.0332521	1.01	0.95-1.08	0.758
Alkaline phosphatase (U/L)	0.0057762	0.0013348	1.01	1.00-1.01	<0.001
Alkaline phosphatase2	−2.00e-06	1.05e-06	0.99	0.9999959-1.0	0.057
ALT (U/L)	0.0099102	0.0025566	1.01	1.00-1.01	<0.001
ALT2	−6.44e-06	3.48e-06	0.99	0.9999867-1.0	0.064
Intercept	−8.735021	0.0001462

^*The formula of the resulting prediction model is:

P_{3-year probability for pancreatic cancer following diabetes diagnosis} = e(Xβ)/1+ e(Xβ)

Xβ = 0.045648 x age − 0.0386336 x BMI + 0.4307414 x PPIs − 0.2418048 x total cholesterol + 0.1995238 LDL + 0.0101408 x LDL² + 0.0057762 x alkaline phosphatase − 2.00e-06 alkaline phosphatase² + 0.0099102 x ALT − 6.44e-06 − 8.735021

Figure 2:

Receiver operator curve of the final model

Table 2:

Model performance at different predicted probability cut-offs for definitive PDA screening.

Risk Thresholds for recommending definitive PDA screening	Diagnostic performance
>0.1%	Sensitivity	66.53%
	Specificity	54.91%
	PPV	0.26%
>0.5%	Sensitivity	15.68%
	Specificity	97.80%
	PPV	1.24%
>1%	Sensitivity	9.32%
	Specificity	99.67%
	PPV	4.7%
>5%	Sensitivity	5.93%
	Specificity	99.96%
	PPV	20.89%
>10%	Sensitivity	2.97%
	Specificity	99.98%
	PPV	20.0%

PPV= positive predictive value

Discussion

We developed a clinical model for the prediction of 3-year risk for PDA in newly-diagnosed prediabetics. The model consists solely of clinical and laboratory information routinely available in general clinical practice. It has good discrimination and calibration. Internal validation using a bootstrapping procedure demonstrated negligible optimism.

Among the patients who developed PDA in our study cohort, the median time from their initial IFG detection to clinical diagnosis of PDA was nearly 1 year. In our previous study among patients with newly diagnosed DM, the corresponding median time from the initial DM diagnosis to PDA diagnosis was about 6 months.⁵ This extended timeframe is highly relevant clinically, as recent data suggest that stage I PDAs progress to stage IV PDAs in about a year.²⁴ Therefore, the current model could substantially extend our ability to detect subclinical PDAs at an earlier stage when the tumor may potentially be more amenable to curative resection. We are aware of only one other PDA prediction model designed to be used among patients with biochemical onset of diabetes.²⁵ However, that model can only be applied among a relatively small fraction of diabetes patients with blood glucose (i.e., fasting blood glucose or hemoglobin A1c) and weight measurements 1 year prior to the diabetes diagnosis. In contrast, our model, which relies only on information readily available at the time of prediabetes diagnosis, can be applied to a much larger proportion of diabetes patients. Except for diabetes medications, which are seldomly used at the time of prediabetes diagnosis, the final list of predictors for pancreatic cancer we identified among those with prediabetes (e.g., age, BMI, PPI use, cholesterol, liver function tests) largely overlaps with the one we identified among patients with new-onset diabetes.⁵ This would suggest that the link between prediabetes and pancreatic cancer is likely mediated by the same mechanisms as the link between new-onset diabetes and pancreatic cancer. Consistent with our prior PDA prediction model developed for new-onset diabetes,⁵ we used a more inclusive minimum age cut-off (i.e., 35 years) to define the study cohort. It is difficult to predict how choosing a higher minimum age cutoff (e.g., 50 years) would affect the prediction model, but the resulting model in that case would not be applicable to the substantial population diagnosed with prediabetes younger than that cutoff.

Our study has limitations. There were missing data in several of the laboratory-based predictors. As a result, we were unable to include some of the potentially relevant variables that might have enhanced the performance of the model. We applied multiple imputation to several other predictors with a lesser extent of missingness. Additionally, the model was developed based on a population-representative UK database. Performance characteristics of the model may differ if applied to the other populations due to potential population differences in the prevalence and management of pre-diabetes and diabetes.²⁶ These issues should be investigated in future retrospective and prospective external validation of our model. Finally, the performance characteristics of our model assume that available imaging modalities will be able to detect smaller, subclinical PDAs at the time of presentation with IFG. Further prospective cohort studies and/or clinical trials will be necessary to test the validity of this assumption.

Currently there is no pancreatic cancer screening recommendation for the general population except in the small group of people with familial or genetic mutations associated with pancreatic cancer. The main reason is that population-wide screening is not efficient due to the relatively low incidence and prevalence of pancreatic cancer as well as the considerable cost and risk associated with the available cancer detection modalities (e.g., MRI/MRCP and endoscopic ultrasound/fine needle aspirate). A risk-tailored early detection approach is likely more feasible. The new-onset diabetic/prediabetic population could be a target population for pancreatic cancer screening. However, the vast majority of new-onset diabetes are not associated with pancreatic cancer, making screening all patients with new-onset diabetes less cost-effective. Our freely-available clinical prediction model provides a safe and easily applicable approach to achieving further cancer risk enrichment within the new-onset prediabetes population, thereby allowing the definitive early detection effort (e.g., EUS, MRI) to be focused on the prediabetic population who are deemed at high risk for subclinical pancreatic cancer. Such a novel, sequential and risk-tailored pancreatic cancer early detection strategy (Figure 3) maximizes efficiency and more importantly may theoretically lead to improved survival for a substantial proportion of pancreatic cancer cases. Using a risk threshold of >0.1% to define high-risk group, up to 66% of all PDA cases diagnosed within 3 years of prediabetes onset can be detected early (Table 2). The number needed to screen for this cutoff would be 380 individuals in order to detect one PDA case. Assuming an average Medicare cost for abdominal MRI/MRCP is $659, the cost of such an approach is expected to be around $250,420 dollars per case of PDAC potential diagnosed about 1 year before clinical diagnosis. Of note, the low specificity of such an approach (56%) is being overcome by the sequential approach of the screening since a definitive diagnostic test would be used among all individuals with high risk according to the model. Because the current analysis was conducted in a setting without any PDA screening, all the PDA cases were diagnosed clinically with the majority being advanced stage cancers. Therefore, dedicated future prospective studies are necessary to evaluate how effectively implementing this sequential risk-tailored screening strategy can shift the stage distribution of PDAs toward early stages.

Figure 3.

Risk-tailored Sequential PDA Screening Strategy in New-onset PreDM

In summary, we developed and internally validated a novel and easily applicable clinical model capable of differentiating patients at high-risk for pancreatic cancer-induced IFG from those with non-cancer related pre-diabetes. If the model can be externally validated, it could be easily applied to all individuals with newly diagnosed prediabetes in order to measure their 3-year predicted PDA risk, and potentially allow early detection of subclinical PDA among those deemed as high-risk.