A Clinical Calculator of Conditional Survival Estimates for Resected and Unresected Pancreatic Cancer Survivors

Abstract

Background

Standard estimates of survival do not reflect temporal changes in risk and are applicable only at the initial diagnosis of cancer. Conditional survival estimates provide more accurate prognostic information for cancer survivors. We sought to calculate conditional survival estimates for patients with pancreatic adenocarcinoma.

Study Design

We identified patients with pancreatic adenocarcinoma diagnosed between 1988 and 2005 in the Surveillance Epidemiology End Results (SEER) cancer registry. We constructed separate multivariate survival models, adjusted for seven clinicopathologic factors, for patients who did and did not undergo radical surgical resection.

Results

Conditional survival probabilities increased over time for all patients with pancreatic cancer regardless of patient characteristics, disease stage, or treatment. For patients with resected stage I, II and III disease, 3-year conditional cancer-specific survival increased from 38% to 70%, 19% to 54%, and 8% to 39%, respectively, over the three years following diagnosis. The relative improvement in survival over time was larger for patients with advanced disease. A customizable, internet browser-based clinical calculator was implemented that may be used to compute, in real time, personalized conditional survival estimates based on an individual’s unique clinicopathologic profile.

Conclusions

Conditional survival estimates provide a more accurate—and typically more optimistic—assessment of prognosis for patients with pancreatic cancer than do traditional survival estimates that apply only at the initial diagnosis.

Pancreatic ductal adenocarcinoma (PDAC) represents the 4^th-leading cause of cancer-related death in the United States.¹ PDAC has aggressive tumor biology and high metastatic potential, and less than 20% of patients with newly diagnosed PDAC survive 1 year following the initial diagnosis. Only 4% survive 5 years.² These statistics often contribute to a sense of nihilism among both pancreatic cancer patients and their medical practitioners.

It is increasingly recognized, however, that the probability of future survival of patients with PDAC is determined by multiple clinical factors, including disease stage at presentation and the use of surgical resection. Even long-term survival is no longer uncommon following multimodality therapy for localized cancer.^3-6 Furthermore, because the risk of cancer recurrence and death from PDAC change significantly over time, the probability of a patient’s future survival is also dependent on the duration of time that has elapsed since the initial diagnosis. For cancers associated with a significant rate of early death, traditional survival estimates applicable to patients with a new cancer diagnosis rapidly become inaccurate and irrelevant. The median or 5-year overall survival estimates that are routinely reported therefore cannot be directly applied to pancreatic cancer survivors. To the extent that personalized survival probability estimates have a critical influence on personal decision-making, disease-related anxiety, and quality of life, the clinical relevance of such estimates depends on the degree to which they reflect not only that patient’s unique clinicopathologic characteristics, but also the time that has elapsed since the diagnosis of cancer.

Conditional survival is defined as the future survival probability or duration that is calculated after a given period of survival.⁷ Unlike standard survival estimates, estimates of conditional survival are time dependent and change significantly over the survivorship period for patients with cancer—particularly those with cancers associated with a low long-term survival probability at diagnosis.⁸ Conditional survival estimates have been reported for patients with various solid cancers, including those of the colon,⁸ rectum,⁹ stomach,¹⁰ skin,¹¹ and breast.¹² In this study, we sought to calculate personalized and clinically relevant conditional survival estimates for individual patients with PDAC. Furthermore, we sought to implement an interactive, browser-based calculator to facilitate clinical application of these data.

PATIENTS AND METHODS

Data Source and Study Population

Data from the Surveillance, Epidemiology and End Results (SEER) 17 program of the National Cancer Institute (released in 2011) were used in this study.¹³ SEER collects cancer incidence, cancer-related variables and survival data from 18 regional registries that cover approximately 26% of the US population.

We evaluated all patients with a histopathologic diagnosis of PDAC as their only malignancy (SEER primary site recodes C250-3 or C257-9 and ICD-O-3 histology codes 8140, 8480, 8481, 8490, or 8500). To allow for at least 3 years of follow-up in December 2008, the patients included in our study were diagnosed from January 1988 through December 2005.

SEER tumor extension codes were used to assign T, N, and M stages according to the American Joint Commission on Cancer (AJCC) 7^th edition algorithm.¹⁴ Tumors were classified as being confined to the pancreas (T1 and T2), extending beyond the pancreas (T3), or extending to the mesenteric vessels (T4). SEER does not provide specific information regarding whether mesenteric vascular involvement is confined to the superior mesenteric or portal vein or involves the superior mesenteric artery. Our approach therefore likely upstaged a small number of patients with isolated involvement of the superior mesenteric vein or portal vein who would be considered T3 (stage II) using the AJCC system but T4 (stage III) using this recoding system. Tumor at distant sites was assessed and coded as metastatic disease (M1). Any lymph node involvement was coded as N1.

We evaluated separately patients who underwent surgical resection and those who did not. We considered patients with non-metastatic disease who had undergone pancreaticoduodenectomy, distal pancreatectomy, or total pancreatectomy to have received a radical oncologic resection.¹⁵ Patients with stage IV cancers who had undergone one of these procedures, as well as all patients who had undergone other non-anatomic operations, were not considered to have had oncologic resection because these operations do not influence survival.¹⁶

Exclusion criteria included age <18 years or >90 years, in situ disease, and lack of histologic or overall survival information. Occurrences also were excluded if the cancer-reporting source was a nursing home, hospice, autopsy, or death certificate; if survival time was <1 month; or if incomplete data precluded AJCC stage assignment.

Statistical Analysis

Survival outcomes for all patients in the study cohort were determined using SEER data through December 2008. Cancer-specific survival was calculated using SEER cause of death recodes. Occurrences were censored if death was due to other causes or if the patient was alive at the last follow-up.

Cancer-specific survival probabilities were estimated by the Kaplan-Meier method. The multiplicative law of probability was used to calculate conditional survival probability estimates among patients with a minimum of 3 years of actual follow-up, as previously described.⁸ Conditional survival represents the probability that a patient with cancer will survive an additional number of years (x) given that the patient has already survived a given number of years (y). For example, to compute the x-year conditional survival for patients who have survived y years, the (x + y)-year cancer-specific survival is divided by the y-year cancer-specific survival.

To adjust for the simultaneous effect of multiple variables on survival, we employed multivariate Cox regression analyses. Covariates adjusted in the prediction model were based on factors identified on univariate analysis as well as those we determined to be clinically relevant. These factors were age, gender, race, tumor site, tumor grade, tumor stage, and radiotherapy. We graphically assessed the proportional hazards assumption on the basis of Schoenfeld residuals.¹⁷

Because of the importance of surgical therapy on survival outcomes, adjusted survival functions stratified by the same 7 factors were evaluated separately for patients who had undergone oncologic pancreatic resection and those who had not. Adjusted cancer-specific survival rates were calculated up to 78 months (6.5 years) for patients who had undergone surgery and up to 45 months (3.75 years) for patients who had not. These survival statistics formed the basis for calculating the adjusted conditional survival. For example, conditional survival probabilities stratified by gender were calculated on the basis of the adjusted survival function for men and women and were controlled for the influence from the other covariates in the final model.

Once the final models had been established, we developed a browser-based clinical application that can be used to compute individualized cancer-specific survival and conditional survival probabilities. The calibration of the online application was validated graphically through the goodness-of-fit statistics for Cox proportional hazards models.¹⁸ The agreement between the observed and model-predicted expected risk of cancer-specific death was plotted on the calibration plot for each decile of the survival outcome.

We performed all analyses with Stata MP software version 11.0 (release 2010; Stata, College Station, TX). This study was exempt from review by our institutional review board.

RESULTS

46,515 patients recorded in the SEER database were diagnosed with a first and only primary pancreatic adenocarcinoma between 1988 and 2005. Following exclusions, 37,135 patients were included in the final analysis. Baseline clinicopathologic characteristics of the cohort are reported in Table 1. In all, 5,736 (15%) patients had undergone an oncologic resection and 31,399 (85%) had not. Among patients who had undergone resection, 15% of the tumors were stage I, 78% were stage II, and 7% were stage III. Among the patients with stage I, II, or III tumors, 41%, 44%, and 12%, respectively, had undergone an oncologic resection. Radiotherapy had been administered to a substantially higher percentage of patients who had undergone resection (47%) than those who had not (19%).

Table 1.

Baseline clinicopathologic characteristics of study cohort (N = 37,135), overall and by disease stage

Characteristic	Total		Stage I-III, resected		Stage I-III, unresected, or stage IV
	No.	%	No.	%	No.	%
No.	37,135	100	5,736	15%	31,399	85%
Age at Diagnosis (yr)
<50	3,034	8%	538	9%	2,496	8%
50-75	25,302	68%	4,317	75%	20,985	67%
>75	8,799	24%	881	15%	7,918	25%
Gender
Male	18,923	51%	2,921	51%	16,002	51%
Female	18,212	49%	2,815	49%	15,397	49%
Race
White	30,051	81%	4,774	83%	25,277	81%
Black	4,295	12%	567	10%	3,728	12%
Asian/Pacific Islander	2,789	7%	395	7%	2,394	8%
Tumor Stage*
I	2,061	6%	835	15%	1,226	4%
II	10,132	27%	4,479	78%	5,653	18%
III	3,403	9%	422	7%	2,981	9%
IV	21,539	58%	NA	NA	21,539	69%
Tumor Site
Head	20,538	55%	4,975	87%	15,563	50%
Body and Tail	7,673	21%	706	12%	6,967	22%
Site unknown	8,924	24%	55	1%	8,869	28%
Tumor Grade
1&2	10,475	28%	3,432	60%	7,043	22%
3&4	8,357	23%	1,856	32%	6,501	21%
Unknown	18,303	49%	448	8%	17,855	57%
Radi otherapy
No	28,369	76%	3,049	53%	25,320	81%
Yes	8,766	24%	2,687	47%	6,079	19%

^*, modified from AJCC 7^th edition as per methods; NA, not applicable.

The median follow-up was 58 (range, 35-90) months for patients with resected cancer and 48 (range, 29 – 96) months for patients with unresected cancer. Kaplan-Meier cancer-specific survival plots for patients with PDAC who had and had not undergone surgery, stratified by disease stage and adjusted for age, gender, race, tumor site, tumor grade, and radiotherapy, are presented in Figure 1.

Figure 1.

Kaplan-Meier plots of adjusted cancer-specific survival stratified by disease stage for patients with PDAC diagnosed in 1998-2005 who (A) had and (B) had not undergone an oncologic resection. Note that the survival curves of patients with unresected stage II and III cancers are superimposed.

The results of the separate multivariate survival models are shown in Table 2. The calibration plot indicated good agreement between the observed and model-predicted outcomes (Supplemental Figure 1). For each disease stage, the median cancer-specific survival of patients who had undergone resection was more favorable than that of those who had not. Among patients who had undergone surgery, advanced disease stage and grade, advanced age, and tumor site in the body/tail of the pancreas were all associated with a higher risk of cancer-specific death. Radiotherapy was associated with a lower risk of cancer-specific death. Among patients who had not undergone resection, patients who had been diagnosed with stage IV disease had had the highest risk of cancer-specific death. Advanced age, male gender, black race, advanced disease stage and grade, tumor site in the body/tail, and absence of radiotherapy were additional independent factors associated with a higher risk of cancer-related death.

Table 2.

Results of multivariate cox regression analysis, by disease stage.

Characteristic	Stage I-III, resected		Stage I-III, unresected, or stage IV
	Hazard ratio	95% CI	Hazard ratio	95% CI
Age at Diagnosis (yr)	p<0.01		p<0.01
<50	1	Ref	1	Ref
50-75	1.05	0.95 – 1.16	1.16	1.11 – 1.21
>75	1.20	1.06 – 1.36	1.49	1.42 – 1.56
Gender	p=0.21		p<0.01
Male	1	Ref	1	Ref
Female	0.96	0.91 – 1.02	0.95	0.92 – 0.97
Race	p=0.07		p<0.01
White	1	Ref	1	Ref
Black	1.11	1.01 – 1.23	1.09	1.05 – 1.13
Asian/Pacific Islander	1.06	0.94 – 1.19	0.90	0.86 – 0.94
Tumor Stage *	p<0.01		p<0.01
I	1	Ref	1	Ref
II	1.74	1.59 – 1.91	1.12	1.04 – 1.19
III	2.68	2.35 – 3.07	1.11	1.03 – 1.19
IV	NA	NA	1.68	1.58 – 1.79
Tumor Site	p=0.03		p<0.01
Head	1	Ref	1	Ref
Body/Tail	1.11	1.02 – 1.22	1.10	1.07 – 1.13
Site unknown	1.23	0.90 – 1.67	1.11	1.08 – 1.15
Tumor Grade	p<0.01		p<0.01
1&2	1	Ref	1	Ref
3&4	1.36	1.27 – 1.44	1.31	1.27 – 1.36
Unknown	1.08	0.97 – 1.21	1.12	1.09 – 1.15
Radiotherapy	p<0.01		p<0.01
No	1	Ref	1	Ref
Yes	0.71	0.66 – 0.75	0.71	0.68 – 0.73

^*, modified from AJCC 7^th edition as per methods. NA, not applicable.

Adjusted conditional survival probabilities were calculated separately for patients who had undergone oncologic resection and for those who had not. We developed an interactive browser-based, conditional survival calculator to facilitate the estimation of conditional survival probabilities for individual patients on the basis of clinicopathologic factors and time parameters that may be selected by the operator. Here we report conditional survival using time parameters we felt to be most relevant to each clinical cohort. The adjusted probability of 3 years of future survival stratified by disease stage—given that patients have already survived 6 – 42 months in 6-month intervals—are depicted in Figure 2A for patients who had undergone surgery. For patients who had not undergone this operation, the adjusted probabilities of 1 year of future survival are reported for patients who have survived 3 – 21 months in 3-month intervals (Figure 2B). Conditional survival probabilities increased over time for each disease stage, whether or not surgery had been performed. The relative increase was larger for patients with advanced disease. For example, the probability of 3 years of future survival for patients with resected stage I cancer relatively increased by 85% (from 38% to 70%) between diagnosis and 3 years, whereas that of patients with resected stage III cancer increased 388% (from 8% to 39%).

Figure 2.

A) Three-year conditional, cancer-specific survival by disease stage (adjusted for age, gender, race, tumor grade, tumor site, and radiotherapy) of patients with stage I-III PDAC who had undergone surgical resection. (B) One-year, conditional, cancer-specific survival by stage (adjusted for the same factors) of patients with stage I-III PDAC who had not undergone resection or who had had stage IV disease. The x-axis represents the duration of survival to date (6-month intervals for surgical patients and 3-month intervals for nonsurgical patients).

Our browser-based conditional survival calculator has been implemented at and is available for clinical use at www.mdanderson.org/pancreascalculator.

DISCUSSION

In this study, we identified a clinically significant association between time after diagnosis and conditional cancer-specific survival probability among patients with PDAC. Using data from more than 37,000 cancer patients treated nationwide, we demonstrated that pancreatic cancer-specific survival improves for all patients over time regardless of patient characteristics, disease stage, or treatment. Furthermore, we developed a browser-based calculator that can be used to compute personalized conditional survival estimates based on an individual’s unique clinicopathologic profile. The survival estimates derived from this tool may be recalculated in real time to inform the personal and clinical decision-making of pancreatic cancer survivors and their caregivers.

Prior reports of long-term survival for patients with PDAC have focused on cancer survivors who have lived >5 years after surgical resection and have shown that the relatively infrequent event of 5-year survival is associated with an improved future prognosis.¹⁹ It is increasingly accepted that cancer survivorship begins at the time of diagnosis, however—not at an arbitrarily chosen future milestone.²⁰ We demonstrated here that the future life expectancy of all patients with PDAC continually improves relative to that at diagnosis over time, regardless of disease stage, age, gender, race, tumor grade, tumor site, or treatment. For patients who undergo resection, the absolute improvement in conditional survival over time is dramatic and reflects both the significant risk for early recurrence and the existence of an established subgroup of patients with favorable cancer biology who survive long-term.³ Indeed, the probability of 3 years of future survival improves from 8 – 38% to 39 – 70% for cancer survivors over the first three years. It must be emphasized, however, that we did not observe a plateau in the conditional survival estimates of these patients, reflecting the potential for cancer recurrence and subsequent cancer-specific death we previously reported as occurring as late as 7 years following diagnosis.³

Relative to the group of patients who undergo surgery, far fewer long-term survivors exist among patients who do not undergo resection. Even among patients with advanced stage or inoperable disease, however, the relative increase in conditional survival is significant over periods of time that are clinically relevant to the patient. The adjusted probability of 1 year of future survival of patients with stage IV disease, for example, increases by a relative 73% (11% to 19%) over the first year (see Figure 2B). Real-time conditional survival estimates therefore provide a more accurate assessment of prognosis for all pancreatic cancer patients than do the routinely-cited 5-year survival rates or median survival statistics that apply only at the time of initial diagnosis of cancer.

The clinical importance of personalized prognostication is well established. Indeed, considerable effort has been expended in the creation of risk nomograms with which the survival probability of individual patients with pancreatic cancer can be calculated on the basis of clinicopathologic variables.^{21, 22} Existing survival nomograms are inexpensive and easy to use but were constructed using data from highly selected patients treated at single, specialized centers and therefore may not be generalizable to larger populations. Moreover, they can be applied only to patients with early stage cancer immediately following surgical resection and do not reflect the continuous changes in survival probability that we have demonstrated among pancreatic cancer survivors over time. For these reasons, those clinical tools have limited clinical utility to most patients with pancreatic cancer.

Our online conditional survival calculator addresses these limitations. This clinical calculator is applicable to all patients with PDAC, regardless of disease stage or treatment. It accurately reflects the simultaneous effect of multiple clinicopathologic factors on survival, and it may be used at any point in time following diagnosis. Furthermore, our calculator allows the user to input the desired duration of time for which future survival probability is computed—an important consideration given the limited relevance of the 3- or 5-year overall survival estimates traditionally reported for patients with PDAC.^{8, 9} Indeed, customization of both time parameters and critical clinicopathologic variables allows clinicians to calculate, in real time, the conditional survival probability estimate specific to a unique patient and relevant to a specific clinical scenario. These estimates can be updated at each surveillance visit and may facilitate personalized decision-making. For example, the personal, economic, and treatment-related decisions made by a young white man with well-differentiated, stage II disease following resection and radiation (who has a 31% chance of surviving 3 years) might differ considerably from those of the same patient who is still alive 3 years later (who has approximately a 65% chance of surviving 3 additional years). Appreciation of improvements in conditional survival over time may, over time, reduce the anxiety and fear of disease progression that exert a negative impact on the quality of life of cancer survivors and their caregivers.^23-27

Our analysis has important strengths and limitations. We evaluated data from a population-based representation of patients with all stages of PDAC. This is a distinction from previously developed tools based on data generated by single institutions with highly selected patients. Although the observed experience of individual centers may differ slightly, our analysis is applicable to the U.S. population of patients diagnosed with PDAC. Furthermore, because of the unselected nature of the study cohort, we were able to develop separate survival models for patients who had and had not undergone resection, which is a critical determinant of long-term outcome. This bears emphasizing given that a minority of patients with PDAC undergo surgery—even those with AJCC stage I or II disease who would otherwise be considered surgical candidates on the basis of tumor anatomy.^28-30

However, some relevant data, including whether systemic chemotherapy was administered, are not available in SEER. This may have influenced our model because chemotherapy may prolong overall survival both as adjuvant therapy for patients with resected cancer and as primary therapy to patients with advanced disease.^{31, 32} Interestingly, we did identify a survival advantage associated with the administration of radiotherapy, even among patients who underwent surgery. Although adjuvant chemoradiation is well-accepted in many American high-volume pancreatic centers, its benefit has not been clearly demonstrated in randomized clinical trials.^32-36 A second limitation is that clinical stage assignment may be difficult to determine in the absence of a resected pathologic specimen and may affect the stage-dependent estimates for patients who did not undergo cancer-directed surgery. However, the use of cancer-directed surgery was a more important determinant of outcome, and in fact the survival estimates for unresected localized tumors were similar to each other. By developing separate models for the resected and unresected cohorts, our analysis separately informs prognosis for patients who were unable to undergo resection and is applicable to the “real-world” scenario of caring for patients with best available clinical stage assignment. The performance of our model might be improved by the addition of clinical variables not currently available within the SEER database, such as performance status, which is known to be of clinical significance for patients with any stage of PDAC,³⁷ or more specific anatomic descriptions of the relationship between the tumor and the superior mesenteric artery and vein. However, as performance status is an important determinant of surgical resection candidacy, stratifying by resection status can be expected to minimize this effect.

In summary, we have shown that conditional survival estimates are a more useful measure of future survival for pancreatic cancer survivors than traditionally-reported survival estimates which are applicable only at diagnosis. We have further shown that the survival expectancy of all patients with PDAC improves over time, regardless of patient characteristics or the pathologic variables associated with the primary cancer. Finally, we have developed a powerful, interactive clinical calculator for generating clinically relevant estimates of future survival that are individualized to patients with PDAC. These estimates may have significant effects on the quality of life of pancreatic cancer survivors.