Markov processes for management of branch-duct intraductal papillary mucinous neoplasms (BD-IPMN) with worrisome features. A, Early resection. B, Long-term surveillance. “M” marks entry point of Markov process. DGE indicates delayed gastric emptying.
State distributions for surveillance (orange) and early resection (blue). Early resection is expected to provide greater cancer-free survival beyond 4.25 years, and greater overall survival beyond 6.5 years.
Projected utility for surveillance (grey) and early resection (blue) via probabilistic sensitivity analysis using 1000 Monte Carlo simulations. Early resection yielded greater expected utility (x-axis) than surveillance in 94% of simulations. dQALY indicates discounted quality-adjusted life-years.
Customize your JAMA Network experience by selecting one or more topics from the list below.
Hu Y, Johnston LE, Shami VM, et al. Comparative Effectiveness of Resection vs Surveillance for Pancreatic Branch Duct Intraductal Papillary Mucinous Neoplasms With Worrisome Features. JAMA Surg. 2018;153(3):225–232. doi:10.1001/jamasurg.2017.4587
For patients with pancreatic branch duct intraductal papillary mucinous neoplasms with worrisome features, does early resection or long-term surveillance result in greater overall utility?
In this comparative effectiveness study, based on a comparative effectiveness Markov model, early resection offered greater overall utility than long-term surveillance for branch duct intraductal papillary mucinous neoplasms with worrisome features. Selection for surgery depends on specific patient, surgeon, and neoplasm characteristics.
Early resection should be considered for branch duct intraductal papillary mucinous neoplasms with worrisome features.
The 2012 international consensus guidelines defined a subcategory of pancreatic branch duct intraductal papillary mucinous neoplasms with “worrisome features,” which may be followed with close surveillance. However, given the poor prognosis of invasive malignancy, the role of early, upfront resection requires further investigation.
To compare the utility of upfront resection vs long-term surveillance. We hypothesized that surveillance of these cystic neoplasms would offer greater long-term utility.
Design, Setting, and Participants
A Markov decision analysis model was constructed to estimate and compare 2 management strategies: early resection and long-term surveillance. Estimates for the utility of outcomes, probabilities of transitions between disease states, and probabilities of surgical morbidity were derived from a literature review of articles published between 1997 and 2014. The comparative effectiveness model’s variable estimates were based on data published predominantly by high-volume, tertiary referral centers. Model probability variable estimates were derived from large, retrospective, single-institution reports. For utility variables, estimates derived from studies using standard-gamble or time-tradeoff methods were given greater weight.
Main Outcomes and Measures
Expected utility was measured in terms of quality-adjusted life years using 3% annual discount. Probabilistic and 1-way sensitivity analyses were performed to assess the potential effects of uncertainty in estimates of key model variables.
Early resection yielded 11.63 quality-adjusted life-years during 20 years of follow-up compared with 11.06 for surveillance. Probabilistic sensitivity analysis indicated that resection has a 94% likelihood of being more effective than surveillance. Early resection obtained greater utility only if each of the following criteria are met: life expectancy is at least 18 years, surgical mortality is less than 4.3%, and baseline preoperative utility is at least 0.78. Additional drivers of the model outcomes include the rate of progression from worrisome to high-risk features and the likelihood of finding cancer on resection for neoplasms with high-risk stigmata.
Conclusions and Relevance
Early resection compares favorably with surveillance in the management of branch duct intraductal papillary mucinous neoplasms with worrisome features. However, careful consideration of patient factors and surgeon outcomes is imperative.
Pancreatic cancer is the fourth-leading cause of cancer mortality in the United States among both men and women, accounting for 40 000 deaths annually.1 The pathway of progression from intraductal papillary mucinous neoplasms (IPMN) to invasive carcinoma accounts for approximately 20% to 30% of pancreatic cancers.2 Presenting at a mean age of 65 years with a slight male predominance,3 these mucin-producing epithelial neoplasms involve the main and/or branch ducts of the pancreas and frequently result in ductal dilatation. Most are asymptomatic on discovery, and with advancements in cross-sectional imaging, incidence has slowly increased over time to more than 2 cases per 100 000 person-years.4,5
Risk of malignancy for IPMN varies by morphology. Those involving the main pancreatic duct have a 56% to 87% risk of malignancy, while branch duct variants (BD-IPMN) have a 6% to 46% risk.6-10 Given this disparity, consensus guidelines recommend early resection for all main duct IPMN, while treatment for BD-IPMN remains controversial. Contributing to the optimal management of BD-IPMN are clinical factors, morphologic features, and patient age. Lévy et al11 reported that the 5-year actuarial risk of developing high-grade dysplasia or invasive cancer in BD-IPMN is 15%. 2012 International consensus guidelines subcategorize BD-IPMN into those with low risk, those with “worrisome features,” and those with high-risk stigmata.12 High-risk stigmata are associated with a 25% risk of invasive malignancy, and therefore, early resection is recommended.13-15 Those BD-IPMNs without any worrisome features have a low likelihood of progression and are recommended for long-term surveillance. However, the optimal management for BD-IPMN with worrisome features remains unclear. These radiographic features include cyst size of at least 3 cm; thickened, enhancing cyst walls; dilation of the main duct to 5 mm to 9 mm; nonenhancing mural nodules; main duct caliber change with distal atrophy; and lymphadenopathy.
Owing to its relatively low incidence and challenging longitudinal follow-up, the level of evidence for BD-IPMN outcomes remains low. Alternative long-term management strategies of BD-IPMN with worrisome features can be compared using Markov process decision analysis models to estimate the likelihood of key events. The objective of this study was to compare estimated outcomes for alternative management strategies for BD-IPMN with worrisome features and to determine clinical utility of resection vs surveillance for this patient population.
Markov process decision analysis models are used to estimate and compare the outcomes of 2 or more interventions through mathematical simulations in which patients in different intervention groups transition to alternative health states over time.16 Follow-up time is incremented by Markov cycles (in this study, 3 months), with each cycle characterized by time spent in a specific disease state (for example, “worrisome BD-IPMN”). Transition probabilities determine the likelihood of moving from 1 disease state to another. For example, the transition probability “high-risk progression” determines the likelihood of moving from the worrisome BD-IPMN disease state to the high-risk BD-IPMN disease state during the period of 1 Markov cycle. Time spent in each disease state is associated with a specified utility value. Estimates for the probability of transitioning from a given health state and utilities specified for each health state are derived from published literature. The overall expected effectiveness of each management strategy is calculated using the accrued utilities for time spent in each Markov cycle over the total period of follow-up.
We created a Markov decision model to represent the patient’s utility perspective and to compare aggressive surveillance against upfront pancreaticoduodenectomy. Major short-term complications of surgery, including pancreatic leak and delayed gastric emptying, are represented in the model. Following initial intervention, the Markov model simulates a 20-years period of follow-up and allows patients to progress through different health states. Costs were not included as outcomes in the model. Because utility in these types of analyses is derived from the patient’s perspective (quality-adjusted life-years [QALYs]), costs must also be analyzed from the same perspective. A cost reference that is frequently used is Medicare reimbursement rates. However, this represents a societal, and not the patient’s, cost perspective. To assess the feasibility of incorporating patient-level costs, we discussed cost with a few select patients who underwent pancreaticoduodenectomy at the University of Virginia School of Medicine, Charlottesville. All patients who participated in this discussion had Medicare insurance coverage, and out-of-pocket expenses for these patients was $0. As such, the decision was made to focus only on comparative effectiveness in this study. This study was reviewed and approved by the University of Virginia institutional review board under a waiver of consent.
The conceptual format of the models for the 2 treatment strategies is summarized in Figure 1. Baseline estimates of survival for 65-year-old men without cancer were obtained from the Centers for Disease Control and Prevention Life Tables.17 In the model, patients in the early resection arm (Figure 1A) undergo pancreaticoduodenectomy upfront for BD-IPMN with worrisome features. Possible health states following resection include no complication, complication (pancreatic leak or delayed gastric emptying), and perioperative death. In the model, the occurrence of major complications is associated with a transient disutility assigned over 1 Markov cycle. Resection may reveal occult invasive carcinoma, with or without regional lymph node involvement, and these events influence both long-term prognosis and adjuvant therapy. Probabilities of subsequent survival and disease recurrence for patients in the alternative health states were estimated from data provided in published studies and are summarized in the Table.13-15,17-40 Per consensus guidelines, patients within the surveillance arm underwent pancreatic imaging.12 While surveillance of patients who had resection for BD-IPMN is imperative given known risk of de novo adenocarcinoma in the pancreatic remnant, to date, there is no consensus on the best surveillance strategy for patients with worrisome-feature BD-IPMN who have undergone resection. Thus, an assumption was made that the long-term utility of resected, disease-free patients would be equivalent to patients in the surveillance arm who have not experienced disease progression. The implication of this assumption is that, across the 2 treatment strategies, any differences in long-term surveillance protocol would not have an effect on long-term utility.
During each cycle of the Markov process, patients with BD-IPMN in each disease state may remain stable, progress to high-risk stigmata, progress to invasive cancer, or transition to death of noncancer causes. In the model, patients who develop disease progression proceed to surgical resection via pancreaticoduodenectomy. Clinical outcomes and adjuvant therapies following resection for disease progression depend on the estimated probabilities of the pathologic stage of the resected specimen.
The baseline clinical scenario for our study was a 65-year-old patient with a noncancer life expectancy of approximately 20 years with incidentally discovered BD-IPMN.17 It was assumed that the patient would be compliant with long-term surveillance and was seeking treatment at a high-volume center experienced in both radiologic interpretation of surveillance imaging and complex pancreatic surgery. Applicability of the model outcomes to older and younger patient populations was explored via sensitivity analysis.
Patients may move from 1 disease state to another following each Markov cycle. Health state transitions are defined by the event probabilities that are summarized in the Table. All model variable estimates were derived from data reported within peer-reviewed literature published within the last 10 years and are summarized in the Table. Large, single- and multiple-institution longitudinal observation studies were the primary source of the probability estimates. Estimates obtained using more than 1 source were calculated using the mean value, weighted by sample size.
The model’s outcome, effectiveness, was measured using discounted QALYs (dQALY), defined as sum of utilities across all Markov cycles with 3% annual discount. For every disease state represented in the model, an estimated utility value is assigned. The accumulation of utility estimates for all disease states within the model across the 20 years of projected follow-up provides the number of QALY attained for each of the model’s treatment arms. All utility variables were discounted at an annual rate of 3% during the period of follow-up. These estimated utilities were also derived from peer-reviewed literature. Where available, utility estimates derived from standard gamble or time-tradeoff methods were given preference37; however, most published utility values were derived from expert consensus or semiquantitative questionnaires. If a disease state’s utility had not been established within peer-reviewed literature, investigator approximations were used. We accounted for the potential inaccuracy of these approximations by selecting wide distributions for sensitivity analysis.
For uncomplicated pancreaticoduodenectomy, a transient disutility was assigned over 1 Markov cycle to represent the routine postoperative recovery period. Major complications, such as pancreatic leak and delayed gastric emptying, were assigned additional disutility values over the recovery period. A separate subgroup analysis was performed to estimate model outcomes for patients with pancreatic body and tail BD-IPMN treated with distal pancreatectomy that accounts for different morbidity and mortality associated with this operation. Surgical outcomes associated with both operations are summarized in the Table. The model assumes that long-term quality of life with chronic surveillance for BD-IPMN and following pancreatic resection for noninvasive disease approximates that of the age-adjusted general population.36,41
Sensitivity analyses were conducted to assess the potential effects of uncertainty in the baseline model estimates for all included components. Sensitivity analyses were performed, allowing each baseline estimate to range between 50% and 200% of the original variable estimate value. One-way sensitivity analyses were performed for all variables to determine which variable(s) had the greatest potential to affect the model’s outcomes. Probabilistic sensitivity analysis was conducted using all model estimates simultaneously, with values for each model estimate drawn at random from a triangular distribution centered on the baseline estimate and bounded by 50% and 200% of the baseline. Monte Carlo simulation was performed to assess the effect of imprecision across all model estimates using 1000 reestimated comparisons of the complete model results, obtained using sets of model estimates drawn at random from the triangular distributions. The results of the probabilistic sensitivity analysis assess the overall effect of uncertainty within the comparison between early resection and long term surveillance. All analyses were performed using TreeAge Pro 2014 (TreeAge Software Inc).
For the baseline scenario of a 65-year-old patient, early resection yielded 11.63 dQALY, compared with 11.06 dQALY for surveillance. The utility advantage of early resection is derived from the difference in long-term clinical outcomes. Figure 2 plots the distribution of time spent in various disease states by intervention arm. Early resection is estimated to outperform close surveillance in cancer-free survival beginning at 4.25 years of follow-up. Similarly, early resection offers an overall survival advantage over surveillance beyond 6.5 years of follow-up. Markov model outcome estimates for distal pancreatectomy rather than pancreaticoduodenectomy were not significantly different. Within this subgroup, the baseline utilities of each strategy were 11.71 dQALY for early resection and 11.07 dQALY for surveillance.
One-way sensitivity analyses were performed to assess the potential effect of uncertainty in each baseline variable estimate of the model’s estimated outcomes for each alternative strategy. These results demonstrate that the model results are contingent on 5 conditions: (1) noncancer patient life expectancy of at least 18 years, (2) perioperative mortality less than 4.3%, (3) annual rate of progression from worrisome features to high-risk stigmata greater than 4%, (4) prevalence of invasive cancer on resection for high-risk stigmata greater than 20.6%, and (5) baseline preoperative utility greater than 0.775. Only when all 5 conditions are satisfied is upfront surgery expected to outperform long-term surveillance for BD-IPMN with worrisome features.
Probabilistic sensitivity analysis was performed with 1000 Monte Carlo simulations using variable parameters drawn at random from triangular distributions of plausible values. The simulated results for the 2 management strategies overlap only over a small portion of the effectiveness distribution (Figure 3), with surgery outperforming surveillance in 94% of the simulations. These results indicate that the principal finding is relatively robust to plausible variation in the values of the model’s component estimates.
Our study demonstrates that, under appropriate conditions, upfront resection for BD-IPMN with worrisome features may be more effective than aggressive surveillance. This advantage is primarily driven by expected improvements in long-term cancer-free and overall survival. By pursuing resection early for pancreatic cystic neoplasms with worrisome features, patients and clinicians may circumvent the substantially higher rate of invasive disease associated with cystic neoplasms with high-risk stigmata. Nevertheless, the model cannot capture patient preference; some patients may find that pancreatic resection represents an insurmountable disutility, and that the possibility of a long-term payoff does not outweigh the near-term morbidity after resection.
Just as relevant as the model’s primary outcome is the set of requisite conditions identified through sensitivity analysis that must be met for upfront surgery to be advantageous. All of these conditions are not surprising, but delineating specific thresholds provides a useful framework. For example, it is intuitive that only patients who have a good life expectancy will glean the long-term benefit of early resection. Our data suggest that good life expectancy entails a noncancer expected survival of at least 18 years. Based on 2010 US life tables published in 2014, a 65-year-old patient’s life expectancy is 19.1 years17; therefore, being younger than 65 years may be a reasonable cutoff for early resection. Similarly, the advantage of early resection is maximized when surgery is performed by surgeons with excellent perioperative outcomes. A 2015 ACS NSQIP study demonstrated 2% perioperative 30-day mortality after pancreaticoduodenectomy for patients with premalignant pancreatic neoplasms.42 However, review of NCDB data among patients with malignant indication for resection demonstrated increase in mortality from 3.7% at 30 days to 7.4% at 90 days following surgery.43 Our sensitivity analysis indicates that those surgeons with perioperative mortality rates higher than 4% should not offer pancreaticoduodenectomy for BD-IPMN with worrisome features.
The 2012 consensus guidelines broadly recommend early resection for BD-IPMN with high-risk stigmata and surveillance for BD-IPMN with worrisome features. Our findings do not necessarily contradict these recommendations. For example, these guidelines indicate that patients with BD-IPMN with worrisome features who are younger than 65 years and those with cysts greater than 2 cm in size may be candidates for resection.12 The first condition concurs with our finding that a minimum life expectancy of 18 years is necessary to consider upfront surgery. The second condition implies that larger cysts may possess a higher likelihood of malignancy or higher rate of progression. Our sensitivity analysis corroborates these recommendations, indicating that those BD-IPMN that are anticipated to progress to high-risk stigmata at a rate of at least 4% per year should be considered for resection. Our baseline estimate for likelihood of finding invasive disease on resection for BD-IPMN with worrisome features, based on data published within the last 10 years, is 10.6%.18,19,24 However, to date, data comparing contributions of individual components within the worrisome-features cyst characteristics are lacking. As data arise that further expand the significance of individual cyst features (such as mural nodules and size), the progression rate and malignancy risk estimates used in this model may require further refinement. Our sensitivity analyses stress these 2 specific data points as critical for further scrutiny. Ideally, pancreatobiliary gastroenterologists and surgeons would be able to draw on a user-friendly comparative effectiveness model for BD-IPMN, use the anticipated malignancy risk of a given cystic neoplasm based on nomograms such as that published by Attiyeh et al,44 and calculate the relative utility of surveillance vs surgery for any given patient.
Ours is not the first study to assess comparative effectiveness in the management of BD-IPMN. In 2010, Huang et al41 performed a cost-effectiveness analysis encompassing all BD-IPMN. They concluded that, while surgery is the most effective option, it is prohibitively costly.41 Their study used a comprehensive societal perspective of cost, taking into account Medicare reimbursement as well as the opportunity cost of missed work owing to perioperative recovery. However, because utility is derived from a patient-level perspective (QALY) and because it is the surgeon and the patient, not society, who are making the decision regarding surgery vs surveillance, there is an inherent disconnect between cost and utility through these methods. Both utility and cost should be evaluated from a patient-level perspective. Furthermore, because the mean age of patients presenting with BD-IPMN is 65 years, a substantial proportion may not be employed, adding further variability into opportunity cost.
The baseline scenario used in this model provides a framework on which many adaptations can be performed. Applying our findings requires consideration of how each specific case differs from the baseline estimates. For example, if a 75-year-old patient presents with BD-IPMN with worrisome features to an institution with perioperative mortality of 3.5%, the relatively short noncancer life expectancy and relatively high perioperative mortality rate should prompt consideration for long-term surveillance. On the other hand, if an otherwise healthy 60-year-old patient presents to a hospital with perioperative mortality of 1%, the potential utility advantage of upfront surgery may be substantially higher. The 5 conditions identified during sensitivity analyses should be used to structure an informed patient-physician conversation weighing the risks and benefits of each strategy. One key patient-related variable that is not well reported in the literature is patients’ long-term utility following pancreatic surgery. While most patients who undergo pancreaticoduodenectomy for benign disease return to their preoperative functional status, a small portion may experience chronic disutility that can range from minor digestive changes to the need for long-term assisted living. Long-term utilities are in need of further exploration and have the potential of changing the results of any comparative-effectiveness study.
This study has several limitations. First, because long-term data regarding the progression patterns of BD-IPMN beyond 5 to 10 years is limited, long-term baseline estimate of a 5% annual rate of progression from worrisome features to high-risk stigmata is uncertain. Second, while the literature supports an approximately 25% rate of finding invasive disease on resection for BD-IPMN with high-risk stigmata, owing to lead-time bias, this rate may be lower for high-risk stigmata discovered during surveillance. Third, there is paucity of objective quality-of-life measures for precancer and cancer disease states that facilitates comparative effectiveness research. We made the assumption that noninvasive disease following resection has the same utility as long-term surveillance for worrisome BD-IPMN and that both these values are equivalent to the general population. However, the disutility of a patient undergoing surveillance for a premalignant lesion is unlikely to be zero and may bias the model in favor of surveillance. Despite this potential bias, resection was the favored outcome with greater dQALY.
Early resection for BD-IPMN with worrisome features is an appropriate treatment option for selected patients and clinicians. The key factors to the decision-making process for resection are the patient’s noncancer life expectancy, the surgeon’s operative outcomes, and the likelihood of disease progression for the presenting cystic neoplasm. Future research should focus on data collection for lifelong surveillance in patients with BD-IPMN, predictors of progression of BD-IPMN with worrisome features, and the utility of precancer and cancer disease states.
Corresponding Author: Victor M. Zaydfudim, MD, MPH, PO Box 800709, Department of Surgery, Division of Surgical Oncology, Section of Hepatobiliary and Pancreatic Surgery, University of Virginia, Charlottesville, VA 22908-0679 (email@example.com).
Accepted for Publication: July 10, 2017.
Published Online: November 22, 2017. doi:10.1001/jamasurg.2017.4587
Author Contributions: Drs Hu and Zaydfudim had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
Concept and design: All authors.
Acquisition, analysis, or interpretation of data: Hu, Johnston, Stukenborg, Zaydfudim.
Drafting of the manuscript: Hu, Stukenborg, Zaydfudim.
Critical revision of the manuscript for important intellectual content: All authors.
Statistical analysis: Hu, Johnston, Stukenborg.
Administrative, technical, or material support: Hu, Shami, Zaydfudim.
Supervision: Stukenborg, Zaydfudim.
Other - faculty development, faculty resources for research support: Adams.
Conflict of Interest Disclosures: None reported.
Funding/Support: This work was supported in part by the institutional National Research Service award T32 CA 163177 from the National Cancer Institute (Dr Hu).
Role of the Funder/Sponsor: The funding source had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.