Neoadjuvant therapy versus upfront surgery for potentially resectable pancreatic cancer: A Markov decision analysis

Background Neoadjuvant therapy has emerged as an alternative treatment strategy for potentially resectable pancreatic cancer. In the absence of large randomized controlled trials offering a direct comparison, this study aims to use Markov decision analysis to compare efficacy of traditional surgery first (SF) and neoadjuvant treatment (NAT) pathways for potentially resectable pancreatic cancer. Methods An advanced Markov decision analysis model was constructed to compare SF and NAT pathways for potentially resectable pancreatic cancer. Transition probabilities were calculated from randomized control and Phase II/III trials after comprehensive literature search. Utility outcomes were measured in overall and quality-adjusted life months (QALMs) on an intention-to-treat basis as the primary outcome. Markov cohort analysis of treatment received was the secondary outcome. Model uncertainties were tested with one and two-way deterministic and probabilistic Monte Carlo sensitivity analysis. Results SF gave 23.72 months (18.51 QALMs) versus 20.22 months (16.26 QALMs). Markov Cohort Analysis showed that where all treatment modalities were received NAT gave 35.05 months (29.87 QALMs) versus 30.96 months (24.86QALMs) for R0 resection and 34.08 months (29.87 QALMs) versus 25.85 months (20.72 QALMs) for R1 resection. One-way deterministic sensitivity analysis showed that NAT was superior if the resection rate was greater than 51.04% or below 75.68% in SF pathway. Two-way sensitivity analysis showed that pathway superiority depended on obtaining multimodal treatment in either pathway. Conclusion Whilst NAT is a viable alternative to traditional SF approach, superior pathway selection depends on the individual patient’s likelihood of receiving multimodal treatment in either pathway.


Introduction
Outcomes for pancreatic cancer remain poor despite advances in surgical technique and adjuvant treatment [1,2].Early complete surgical resection is the only potentially curative treatment, with surgery followed by adjuvant therapy becoming the standard of care for resectable pancreatic cancer [3].Up to 50% of patients fail to receive adjuvant therapy due to: post-operative complications, early metastases nullifying the potential benefits of high-risk surgery, and reduced performance status [4,5].Neoadjuvant therapy (NAT) has emerged as an alternative to traditional surgery-first (SF) approach.Postulated benefits include: avoiding futile surgery by identifying aggressive tumour types, eliminating micrometastesis, multimodal treatment completion and increased R0 resection rates [6,7].
There is a lack of randomized controlled trials (RCTs) comparing SF and NAT [8].Despite promising results from cohort studies and phase II trials, meta-analysis have reported only marginal benefit of NAT in terms of overall and disease-free survival [7,[9][10][11][12].Critics have therefore highlighted the limitations of drawing optimistic conclusions from small studies that are underpowered [6,7].Two previous Markov decision-analysis found marginal benefit with NAT for resectable only cases [9,13].Only one previous Markov decision analysis compared efficacy of both pathways for potentially resectable disease and found no conclusively superior pathway [14].
The aim of this study is to synthesize best current evidence within a Markov decision-analysis model comparing SF and NAT pathways for the management of potentially resectable pancreatic cancer on an intention-to-treat basis.This study aims to improve on previous iterations by performing Markov cohort analysis assessing impact of treatment received on survival outcomes.

Markov model
TreeAge Pro 2017 (TreeAge Software Ins., Williamstown, MA) was used to construct a Markov cohort decision analysis model in an advanced decision-tree format comparing base case, surgery first followed by adjuvant therapy (which included chemotherapy, chemoradiotherapy, or both), to NAT (which included chemotherapy and/or chemoradiotherapy) followed by re-staging and, if possible, surgical resection (Fig 1).Upon completion of treatment, cohorts entered the Markov health-state transition model with possible survival states including: alive without disease, alive with disease and dead.Each Markov cycles equated to 1 month with maximum follow-up of 60-cycles or until death.

Outcome measures
Cumulative payoffs were calculated in life months and quality-adjusted-life-months (QALMs), which scaled survival from 0 (equivalent to death) to 1 [9,13] based on indicies taken from The inclusion criteria was RCTs and prospective phase II and III studies of NAT for treatment of pancreatic cancer, published in English language since 2000, involving chemo/radiotherapy-naive human subjects over 18 years of age with preoperatively staged pancreatic cancer as potentially resectable.Included trials had to report: protocol design, number of participants per arm, median age and co-morbidities of subjects, pre-treatment staging of pancreatic cancer, toxicity profile, results of post neoadjuvant re-staging, resection rates, postoperative complications defined by Clavien-Dindo system, and survival data.Retrospective and cohort studies, case series and case reports were excluded as were studies from identical patient cohorts and trials involving intra-operative radiotherapy and trials including disease other than pancreatic cancer.
As the majority of trials were single arm, to populate the SF pathway the same databases were searched for RCTs of surgery and adjuvant therapy, with the same inclusion and data reporting criteria (S2 Fig) .The outcomes of this group could introduce biased because by definition these patients have survived surgery and not developed early metastatic disease and also had to have adequate performance status to be randomized to adjuvant therapy even if they did not receive adjuvant therapy.To overcome this issue cohort studies comparing NAT and SF, with the otherwise same inclusion criteria and data reporting requirements, were also included in the SF arm and solely used to offer comparison across outcomes of resection, R0 resection rates and receipt of adjuvant therapy (S3 Fig) .The Cochrane Collaboration's risk of bias tool [21] and ROBINS-I tool (Risk Of Bias In Non-randomized Studies-of Interventions) [22] were used to assess the quality and risk of bias of each included trial (S1 and S2 Tables and S4 Fig) .Furthermore the potential impact of bias and uncertainty on all variables within the model were extensively tested through probabilistic and deterministic sensitivity analysis.

Statistical analysis
Markov model transition probabilities were based on weighted pooled estimates of proportions from included studies, calculated using Freeman-Tukey arcsine square root transformation under random effects model to account for heterogeneity [23].Survival time was based on time from diagnosis.Gillen et al. approach to calculating weighted median survival time was used as evidence has shown that weighted averaging of medians cannot achieve unbiased pooled estimates of survival time [24,25].This approach is based on averaging parameter estimates of a presumed density function of survival.The pooled distribution parameter is used to recalculate the estimate of the median from the pooled distribution parameter [24].In this case the pooled distribution parameter is the exponential distribution which implies a time constant hazard rate corresponding to the sole distribution parameter λ.From this the weighted estimate of median survival (m p ) is derived from the formula [24]: where m i is median survival within the study population i (with i being 1 to k where k is the number of included studies) [24].w i is the study specific weight function derived from number of study participants divided by total number of evaluable patients [24].

Sensitivity analysis
Model uncertainties for all included components were tested with one and two-way deterministic sensitivity analysis with baseline transition probabilities for each variable altered between highest and lowest reported values.Probabilistic Monte Carlo sensitivity analysis was set to 10000 cycles with model probabilities sampled from the data distribution of each variable.Data for each variable was fitted against 55 possible distributions with best fit determined by the Anderson Darling statistic.

Results
50 phase II/III studies met the inclusion criteria and were included in the NAT arm of the model, 4 of which were randomized.9 of these studies offered comparison with upfront surgery.For the SF pathway 15 studies were RCTs, 10 of which offered comparison between adjuvant regimes, 5 of which offered comparison between adjuvant therapy and surgery only.16 cohort studies were also included in the SF pathway to offer comparison across outcomes of resection rates, R0 resection, and rates of receiving adjuvant therapy (Table 1).Probability estimates and ranges and quality of life utilities are displayed in Table 2.

Deterministic sensitivity analysis
Deterministic sensitivity analysis tested the sensitivity of the results of the model to variations in parameters of specific model variables by altering the parameters between highest and lowest reported values.One-way deterministic sensitivity analysis determined the effect on the overall results of the model by varying the parameter of each variable individually.Two-way deterministic sensitivity analysis determined the effect on the model of altering the parameters of two variables simultaneous.One-way deterministic sensitivity analysis showed that NAT was the superior treatment pathway if the probability of achieving resection in the NAT pathway was greater than 51.04% or the probability of achieving resection in the SF pathway was less than 75.68% (Fig 2).
Two-way deterministic sensitivity analysis demonstrated that treatment superiority depended on receiving multimodal treatment (resection in NAT pathways and adjuvant therapy in SF pathway).

Probabilistic sensitivity analysis
Probabilistic sensitivity analysis tested the level of confidence in the model output in relation to uncertainty in model input by determining the distribution of the input data for each variable from the median, standard deviation and variance of the input data.The input data was then fitted against 55 possible data distributions with the best fit determined by the Anderson Darling statistic (Table 2).All possible parameter values for each variable within the model were therefore tested by drawing probabilities from the data distribution when probabilistic Monte Carlo sensitivity analysis was set to simulate 10000 patients cycling through the model.
The results of probabilistic Monte Carlo sensitivity analysis showed that SF gave a mean survival time of 19.72 months (range 5.57-22.95)compared to 17.16 months (range 16. 50-17.38)for NAT with standard deviation 2.68 and 0.19, and variance 7.17 and 0.04 in SF and NAT pathways respectively.When minimum significant difference was set at 3.65 months or greater, the model reported indifference in superior pathway selection frequency.

Discussion
The role of NAT in treatment of pancreatic cancer is an ongoing area of debate [107].Markov decision analysis is a powerful tool offering analysis of complex medical decisions therefore this study provides important interim analysis and contributes to only three existing studies that utilize evidence-based Markov decision-analysis to compare NAT and SF pathways [9,13,14].Our analysis showed that SF pathway gave an additional 3.5 months (2.25 QALMs) but neither pathway was conclusively superior.Markov cohort analysis of outcomes where multimodal treatment was received in both pathways (resection in NAT and adjuvant therapy in SF) demonstrated superior outcomes with NAT pathway.One-way sensitivity analysis showed that NAT was superior on in intention-to-treat basis if the probability of resection in the NAT pathway was above 51.04%.Base-case probability of resection in NAT pathway was 41%.This could represent NAT pathway allowing time to filter out more aggressive tumours, hence avoiding futile, expensive and high-risk surgery [6,7].Conversely critics of NAT would argue that this indicates loosing the window of resectability [6,7].Critics of NAT have also raised concerns that it could increase post-operative complications, a claim not substantiated in our analysis, which is also corroborated by previous studies [9,14].Three Markov Decision-Analysis studies comparing NAT and SF pathways for pancreatic cancer exist [9,13,14].One of these studies most closely resembled this study by focusing on potentially resectable pancreatic cancer, therefore including borderline and locally advanced cases in the NAT pathway to capture the effect of conversion to resectability on overall pathway analysis [9].As with our findings it did not demonstrate on overall conclusively superior pathway on an intention-to-treat basis (NAT 18.6 months versus 17.1 months) [14].Two other Markov decision analysis studies have reported superiority of NAT pathway but they focus on resectable only cases and based their model on literature from a single search engine [13,14].One study, based on phase I/II trials, reported 22 months for NAT versus 20 months for SF [9] but their analysis mostly included retrospective studies [13].Preliminary results from the PRE-OPANC-1 trial, a multicenter phase III RCT comparing NAT and SF for borderline resectable PC, have reported improved survival with NAT on an intention-to-treat basis (17.1 months versus 13.5 months) [108].Although PREOPANC-1 is a very different study to the one presented here in terms of design and statistical methodology, the results do echo our findings in reporting that higher reported resection rates in the SF pathway do not equate with superior overall survival time for patients treated in a SF pathway [108].Furthermore the subgroup analysis of resected cases in the PREOPANC-1 trial reported superior survival time with NAT (29.9months versus 16.6months), which further corroborates the results our Markov cohort analysis [108].

Strengths and limitations
Improving upon previous iterations, this study, based on a comprehensive search of multiple search engines, went beyond intention-to-treat base-case analysis to perform Markov cohort analysis of treatment received, which provided important results favoring NAT.These findings corroborate those of prospective and retrospective experiences of NAT, which have demonstrated favorable survival outcomes ranging from 26 to 45 months [13] and findings reported by Abbott et al. who, although performing Markov analysis for cost-effectiveness, reported similar outcomes [109].
Considering at base-case 39% in SF pathway did not received adjuvant therapy and 59% in NAT pathway did not undergo resection, this adds an important dimension to the ongoing debate.Two-way sensitivity analysis demonstrated that the superior treatment pathway depended on an individual's probability of receiving multimodal treatment in either pathway.This highlights the need to work towards personalized predictive medicine to assist in the selection of the most appropriate treatment pathway at individual patient level.
Like existing Markov decision analysis based on data from published studies, this study also shares the limitations of the existing body of evidence: heterogeneity, lack of randomization, potential bias, small and underpowered studies [6,7].Furthermore definitions of radiological and surgical resectability, R0 resection and staging protocol can vary across trials further compounding the issue of heterogeneity of synthesized data [8].Such heterogeneity could account for why, at base-case analysis, the probability of R0 resection in the NAT pathway was smaller than anticipated particularly when considering that the PREOPANC-1 trial has reported higher rates of R0 resection (65%) in the NAT arm of their trial [8,108].Although the uncertainty of this variable was extensively tested through sensitivity analysis and found not to affect the overall model outcomes, this highlights the potential impact of heterogeneity of data on model output.To address the issue of heterogeneity, this study based transition probabilities on weighted pooled proportion estimates calculated using Freeman-Tukey arcsine square root transformation under random effects model [23].Furthermore probabilistic Monte Carlo   Unlike previous Markov decision-analysis, weighted survival times were based on the Gillen et al. formulae as evidence has shown that unbiased pooled estimates of median survival times cannot be achieved by weighted averaging of medians [24,25].Quality adjusted survival time is limited by the lack of studies measuring quality-of-life across the treatment trajectory for pancreatic cancer.This study utilized best available data that was shared with existing decisionanalysis studies, which enhanced comparability.Our model did not assume return to full health after an intervention or event but accounted for the impact of therapy, surgery, complications and disease reoccurrence when calculating quality-adjusted survival.For both survival and quality-adjusted survival, uncertainty was rigorously tested across every variable in the model through probabilistic and deterministic sensitivity analysis.

Conclusion
In the absence of large multicenter RCTs, this study utilizes best currently available evidence in a Markov decision analysis model to provide an important interim source of information [13,14] to inform the ongoing debate regarding best treatment for potentially resectable pancreatic cancer.On an intention-to-treat basis conclusive superiority of either pathway could not be concluded.However, Markov cohort analysis demonstrated superiority of NAT pathway if multimodal treatment was received.This highlights three important future directions for research: 1) cost-effectiveness analysis of NAT versus SF pathways 2) assessment of treatment pathways of resectable only cases hence offering a true like-for-like comparison and 3) developing methods of personalized predictive statistical modeling to provide individualized predictions of outcomes across competing treatment pathways to support shared clinical decision-making.

Fig 1 .
Fig 1. Overview of the structure of the Markov decision-tree.https://doi.org/10.1371/journal.pone.0212805.g001 Fig 3A shows the thresholds at which competing pathways offer superior outcomes with Fig 3B providing corresponding probability thresholds and predicted resulting quality-adjusted survival time.

Fig 2 .
Fig 2. One-way deterministic sensitivity analysis.This figure shows the effect of altering the baseline probability of resection first in the NAT pathway then in the SF pathway on overall model outcome.https://doi.org/10.1371/journal.pone.0212805.g002

Fig 3 .
Fig 3. Two-way sensitivity analysis.Y-axis shows probability of receiving adjuvant therapy in SF pathway and x-axis shows probability of receiving resection in NAT pathway.The red area represents where patients, given competing probability of receiving multimodal treatment in competing pathways, would benefit from SF approach.The blue area represents where NAT would be the superior treatment option in terms of quality-adjusted survival.The corresponding predicted survival time in QALMs are detailed below.X and Y-axis provide altering probabilities of multimodal treatment in each pathway with corresponding survival time given in QALMs.https://doi.org/10.1371/journal.pone.0212805.g003

Table 1 . Summary of included trials. Reference Type of Study Treatment Regime N = Disease Free Survival in months Overall Survival in months
[26]ukhun et al.[26]Prospective Phase II Trial CRT + surgery 20 13.4 Cardenes et al.[27] Prospective Phase II Trial CRT + surgery 28 10.3 Casadei et al. [28] Prospective Phase II Trial CRT + surgery 18 28.3Cetin et al. [29] Prospective Phase II Trial CRT + surgery 11