SEARCH

SEARCH BY CITATION

Keywords:

  • competing risk analysis;
  • survival analysis;
  • proportional hazards;
  • prostate cancer;
  • nomograms;
  • prognosis;
  • shared decision-making

Abstract

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. FUNDING SUPPORT
  8. REFERENCES

BACKGROUND:

Population-based cancer registries that include patient follow-up generally provide information regarding net survival (ie, survival associated with the risk of dying of cancer in the absence of competing risks). However, registry data also can be used to calculate survival from cancer in the presence of competing risks, which is more clinically relevant.

METHODS:

Statistical methods were developed to predict the risk of death from cancer and other causes, as well as natural life expectancy if the patient did not have cancer based on a profile of prognostic factors including characteristics of the cancer, demographic factors, and comorbid conditions. The Surveillance, Epidemiology, and End Results (SEER) Program database was used to calculate the risk of dying of cancer. Because the risks of dying of cancer versus other causes are assumed to be independent conditional on the prognostic factors, a wide variety of independent data sources can be used to calculate the risk of death from other causes. Herein, the risk of death from other causes was estimated using SEER and Medicare claims data, and was matched to the closest fitting portion of the US life table to obtain a “health status-adjusted age.”

RESULTS:

A nomogram was developed for prostate cancer as part of a Web-based Cancer Survival Query System that is targeted for use by physicians and patients to obtain information on a patient's prognosis. More nomograms currently are being developed.

CONCLUSIONS:

Nomograms of this type can be used as one tool to assist cancer physicians and their patients to better understand their prognosis and to weigh alternative treatment and palliative strategies. Cancer 2012. © 2012 American Cancer Society.


INTRODUCTION

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. FUNDING SUPPORT
  8. REFERENCES

The Surveillance, Epidemiology, and End Results (SEER) Program is a recognized resource for information concerning cancer patient survival.1 Individuals seeking information regarding prognosis from the National Cancer Institute (NCI)'s Cancer Information Service are routinely referred to the SEER Web site, on which reports are available that include survival statistics by race, sex, stage of disease, age, and cancer site.2 Information regarding survival obtained from cancer statistics reports is generally net survival, and most often estimated using relative survival,3 which indicates the likelihood of dying of causes related to a patient's cancer in the absence of competing risks of death. Because it is not influenced by changes in other-cause mortality (eg, large changes in cardiovascular disease mortality over time), net survival is useful in quantifying temporal changes in the excess mortality experienced by successive calendar-year cohorts of cancer patients as a result of treatment and screening interventions. Although useful in quantifying changes in patient survival, net survival does not reflect the actual survival experience of cancer patients.

When considering the problem of predicting survival for an individual patient, crude survival measures (ie, measures that account for the competing risks of dying of other causes) are of greater relevance. Cronin and Feuer4 proposed methodology for calculating the crude cumulative cause-specific probabilities of death under the assumption of independence among competing risks for a cohort of patients whose risk of other-cause death can be represented by general population life tables. In the current study, we extended the work of Cronin and Feuer4 and proposed crude cumulative measures of the risk of death from cancer and other causes that were made as specific as possible to an individual patient. The methodology proposed herein is used to develop prediction tools or prognostic nomograms that provide crude cumulative cause-specific estimates of the risk of death based on the input of a profile of values of prognostic variables describing a patient and their cancer.5 These nomograms will be part of the Cancer Survival Query System (CSQS), a resource for clinicians and patients. A unique feature of the CSQS nomograms is the use of a “health status-adjusted age (HSAA)” or physiologic age, which may differ from the person's chronologic age. The HSAA is estimated by modeling the risk of other causes of death as a function of the number and type of comorbid conditions using Medicare claims data, and then matching the results for an individual to the closest fitting portion of the US life table. In this article, we described the development of a nomogram for prostate cancer.

MATERIALS AND METHODS

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. FUNDING SUPPORT
  8. REFERENCES

Data Sources

Sources used included cancer survival data from the SEER Program,1 US life tables from the National Center for Health Statistics (including life expectancies by age),6 and Medicare Claims data from the Center for Medicare and Medicaid Services.7 Medicare claims data were linked to SEER data and a 5% sample of Medicare patients without cancer in SEER areas was also used.

SEER Data

The SEER Program includes cancer registries that routinely collect population-based data on cancer site/type, extent of disease (including stage at the time of diagnosis), first course of treatment, and various demographic variables (including age, race/ethnicity, sex, and vital status annually including the cause of death).8 Follow-up was 97.2% complete for prostate cancer.9 Data from the SEER13 registries (minus Alaska)10 covering 20.3% of the total US population in 2005 were used. The coding of extent of disease as performed by SEER is modified periodically to reflect medical practice changes in the staging of cancer.8 To best reflect current staging for prostate cancer, we included only diagnoses made between 1995 and 2005.

Methods

Estimating Crude Cumulative Mortality for an Individual Patient: The Profile Method

We proposed an extension of the work of Cronin and Feuer,4 in which the established theory of competing risks is described along with a methodology to compute crude cumulative mortality for discrete time intervals in a life table setting. We used a discrete time actuarial approach because large registry databases have many tied death and censoring times. As in the study by Cronin and Feuer,4 under an assumption of conditional independence of the risk of dying of cancer and other causes, we estimated the crude cumulative probabilities of death through time interval M from cancer (GcM), and from other causes (GoM) for a cohort as

  • equation image(1)
  • equation image(2)

in which equation image is the life table estimate of the all-cause survival rate in interval i; equation image is the expected survival rate in interval i for the patient group as obtained from US life tables based on the age, race, sex, and calendar year of diagnosis mix of the group; and equation image is the estimated conditional net cause (ie, cancer)-specific survival probability in interval i. Equations (1) and (2) are identical to those presented in Cronin and Feuer,4 except that in their article, equation imageis estimated based on relative survival, whereas in the current study it was based on cause-specific survival using cause of death. Net cancer-specific survival can be estimated using relative survival,3 which uses all-cause survival divided by expected survival from a US life table matched to the age, race, and gender distribution of the patient cohort, or uses cause-specific information based on cause of death. Recently, improved algorithms for determining which specific death codes should be counted as “cancer events” in SEER data have been developed,11, 12 and those algorithms were used in the current study to calculate cause-specific net survival.

Next, we extended the study by Cronin and Feuer4 to the case of a specific cancer patient j with characteristics zj, which influence net cancer-specific survival; and characteristics wj, which influence net other-cause survival. Generally, zj and wj will have some factors in common (eg, age, race, and sex). Equations (1) and (2) can be rewritten as

  • equation image(3)
  • equation image(4)

Because of the independence assumption, which allows all-cause survival equation image to be partitioned into 2 components (net cancer-specific equation image and net other-cause equation image survival), the factors equation image in equation image do not have to be in the registry's database and can be obtained from other sources.

Because zj and wj together represent a profile of prognostic factors for patient j, this formulation is referred to herein as the “Profile Method” for estimating crude cause-specific cumulative mortality. Cox regression for grouped survival data13 can be used to obtain conditional net cancer-specific survival in interval i for patient j (ie, equation image). Parallel methodology for the profile method using relative survival is described in a technical report.14

Variance estimates for equation image and equation image (equations 3 and 4) can be derived by the delta method. Details of estimating these variances are presented in a technical report.14

Estimating Net Other-Cause Survival For an Individual

Net other-cause survival for individual j (ie, equation image) was developed using SEER data linked to Medicare data, and a 5% sample of Medicare data for individuals without cancer residing in SEER areas. Using Cox regression analysis,13 the prognostic significance of the presence or absence of a claim for 15 conditions used as part of the Charlson comorbidity index15 in the year before a diagnosis of cancer for patients aged ≥ 66 years diagnosed from 1992 through 2005 and including selected first-order interactions was determined using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis and procedure codes to identify comorbidities in the SEER-Medicare linked data, in which the event was death from causes other than cancer. A summary comorbidity index was obtained as the weighted sum of Cox regression coefficients for the 15 conditions and used for predicting other-cause survival up to 10 years. Previous studies have demonstrated the usefulness of this approach in developing prognostic comorbidity indices.16, 17 The index was then calculated for both the SEER-Medicare data for all cancers and a 5% sample of Medicare patients without cancer in SEER areas to obtain predicted other-cause–specific survival based on Cox regression analyses stratified by individual age. Variables in this model included sex, race, a continuous function of comorbidity using a restricted cubic spline, and cancer status (cancer vs noncancer, indicating whether the data came from SEER-Medicare or the 5% sample of noncancer patients). Noncancer patients were included in the analysis to increase the available sample size, especially at older ages. The predicted other-cause survival for a specific covariate profile as obtained from these Cox regression analyses was then compared with similar patterns by age in the year 2000 US life tables (the approximate center point of the years of diagnosis included in the analysis) and the closest fit provided the basis for transforming an individual's chronologic age into a HSAA. For example, using this approach, a 67-year-old white unmarried male cancer patient with no comorbidities has a HSAA of 61 years, whereas a white unmarried male patient of the same age with ulcers and diabetes has a HSAA of 73 years. The final estimate of equation image for patient j was obtained using their HSAA in conjunction with the most recent life tables available (ie, 2006) to make the estimated survival used as up to date as possible. Further details of this methodology will be available in a forthcoming publication.

The physician, counselor, or patient may subjectively modify the HSAA further based on comorbid conditions not included in the calculator or other risk factors, and the new HSAA would then be used to obtain the patient's other-cause survival. For individuals aged < 66 years at diagnosis, no calculator was available, and other-cause survival was based on their chronologic age using the 2006 US life table, but the user was allowed to enter a HSAA based on a physician's subjective assessment of the patient's health status. Using the HSAA, the 2006 US life table was also used to compute the patient's natural life expectancy6 if the patient did not have cancer, which is a useful summary measure for assesing fitness for surgery and other treatments.

Estimating Net Cancer-Specific Survival for An Individual Patient

Cause of death information (ie, equations 3 and 4), available in SEER data,11, 12 was used for estimating net cancer-specific survival for patient j. Covariates included in each Cox regression model were those associated with prostate cancer along with age, marital status, year of diagnosis, and, for patients aged ≥ 66 years, comorbid conditions. Analyses were stratified by the major prognostic factors because the large sample sizes allowed fitting for each of these groups separately. Year of diagnosis was modeled as a continuous covariate, and the crude cumulative mortality estimates provided are for the most recent diagnosis year available (2005) to represent the prognosis of a recently diagnosed patient.18 All analyses were performed using 3-month discrete time intervals. Variable selection was not performed because we were primarily interested in predicting survival (rather than identifying statistically significant explanatory variables), and keeping the models consistent across strata.

We believed that age and comorbidity should be modeled using continuous functions rather than step functions based on grouping values of each of these variables because continuous functions better represent the effects of these variables on survival. To determine what continuous functions to use, the data were fit using indicator variables for selected ranges of age and comorbidity scores as well as a restricted cubic spline and a linear function for both variables. Graphs of all of these fits were reviewed and decisions were made by comparing the continuous functions with the nominal indicator function to determine the appropriate continuous function to use for the entire ranges of age and comorbidities.

After these decisions were made, the models were refit to the data in which all data values greater than a specified point or less than a specified point as appropriate were assigned values equal to that for the specified point. Graphs of the effects of age and comorbidity based on the various functions originally considered were again examined for all of the strata, and final decisions were then made regarding the model to use for each strata to compute net cancer-specific survival rates.

Prostate Cancer
Case Selection and Definition of Cause of Death

The net cancer-specific survival information provided for prostate cancer was obtained from patients whose first cancer was an invasive cancer of the prostate diagnosed between 1995 and 2005 with a histologic type of adenocarcinoma (3rd edition of the International Classification of Diseases for Oncology19; histology code 8140) and a behavior code of 3 (ie, malignant). This time period was used because of the availability of the relevant extent of disease information necessary to perform up-to-date staging.

Excluding patients who were alive with no available follow-up and those with only autopsy/death certificate diagnoses, there were 135,158 potentially eligible patients ages 40 to 66 years, and 206,812 patients who were ages 66 to 94 years. In the former group, 112,127 patients (83%) had complete information available regarding all prognostic variables and cause of death, and these patients were used in the Cox regression analyses for this age group. For those patients aged 66 to 94 years, those patients not included in the analysis were those who were not linked to Medicare data (21,432 patients), and those who did not have at least 12 months of Medicare data before the diagnosis of their cancer or who were treated as part of a health maintenance organization (64,599 patients). The remaining patients (120,282) were linked to Medicare data. Of these, 91,419 patients (44% of the original cohort and 76% of those with sufficient linked Medicare data) had complete information available regarding all prognostic variables including cause of death and were used in the Cox regression analyses to obtain net cancer-specific survival for this age group.

Patients aged < 40 years and those aged ≥ 95 years were excluded because of the sparseness of the data for younger patients and the misclassification of cause of death for older patients.11 Cause of death is used in the current study in the survival analysis and death from prostate cancer was coded using recent innovations developed by the SEER Program in the determination of what specific death codes should be treated as “events” for estimating cause-specific survival for various cancers.11, 12

Prognostic Variables

Because of treatment practices for prostate cancer, 3 independent staging systems were developed: “preclinical” for patients who had not yet been treated, “pure clinical” for patients who did not elect to undergo surgery, and “pathological” for patients who were treated surgically. We were able to implement the pretreatment clinical staging system only because SEER maintains information on the clinically determined and pathologically determined extent of disease for prostate cancer. However, clinical lymph node status was not known for patients who were treated surgically and this required an assumption that lymph nodes were clinically negative for these patients in specifying their pretreatment clinical stage.

Prognostic variables for prostate cancer that were available in the SEER database included stage; Gleason score (GS)20; marital status; and the demographic variables age, race, and sex. The staging system used herein is generally preferred by many physicians who treat patients with prostate cancer.21 The definitions of stage of disease are based on the T, N, and M definitions in the sixth edition of the American Joint Committee on Cancer staging manual.22 Two analyses were performed for each of the stages: 1 for patients with comorbidity information (aged ≥ 66 years) and 1 for patients aged < 66 years. The stages of disease noted within each group were localized inapparent, localized apparent, locally advanced, lymph node disease, and distant metastases for all 3 staging systems except for the pathological group, in whom the localized inapparent stage was not applicable (there was only a “localized” stage group for patients in the pathological group) and the distant metastases stage was not included because, for the most part, only patients with clinically localized disease were treated surgically and only a small percentage of these patients tend to be upstaged to distant disease after surgery.

Analyses were stratified based on the major stage groupings, except that the locally advanced group was combined with the lymph node disease group because of the small sample sizes within the lymph node disease group. Cox regression models13 were fit separately for specified strata based on stage (ie, localized inapparent, localized apparent, locally advanced combined with lymph node disease, and distant metastases). Covariates in the model included substages of localized inapparent disease (T1a vs T1b vs T1c), locally advanced disease (T3 vs T4), and lymph node disease (T1-T3 vs T4). Other variables were GS (2-7 vs 8-10), interactions of GS with substages of disease, age (effect modeled as a restricted cubic spline23 but flat for ages 90-94 years), race (white vs black vs other), comorbidity score for patients aged ≥ 66 years (effect modeled as linear but flat for scores of ≥ 1.5), and calendar year of diagnosis (linear). All variables were evaluated graphically for proportionality.

RESULTS

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. FUNDING SUPPORT
  8. REFERENCES

The strata definitions and sample sizes for prostate cancer are given in Table 1 along with the TNM definitions of the stages included in each stratum. A Cox regression model13 was fit for each of the 20 cells in Table 1. An example of a Cox regression fit for 1 stratum is given in Table 2 for patients with preclinical localized inapparent prostate cancer for whom comorbidity data were available. The parameter estimates and their standard errors and associated relative risks are also provided along with P values indicating the statistical significance of the regression coefficients. The Cox regression tables for all of the strata are presented in an NCI technical report.14

Table 1. Strata Definitions and Sample Sizes for Prostate Cancer Cox Regression Modeling
StageTNM DefinitionPretreatment ClinicalPure ClinicalPathological
Aged <66 YearsaAged 66-94 YearsbAged <66 YearsaAged 66-94 YearsbAged <66 YearsaAged 66-94 Yearsb
  • a

    No comorbidity.

  • b

    With comorbidity.

  • c

    For “Pathological,” there is only a “localized” group based on the pathology findings from surgery.

Localized (inapparent)T1a, T1b T1c N0 M049,01934,52719,39525,315
Localized (apparent)T2 N0 M056,97749,31721,82435,48243,984c11,896c
Locally advanced; Nodal diseaseT3-T4 N0 M0; T1-T4 N1 M0333236152064273518,9226002
Distant metastasesAny T, any N, M12910396027413452
Totals 112,23891,41946,02466,98462,90617,898
Table 2. Cox Regression Estimates for Pretreatment, Localized, Inapparent Prostate Cancer for Patients Aged ≥66 Years, Including Comorbidity
VariableBetaSE of β/θRelative RiskP
  • Abbreviation: SE, standard error.

  • a

    Restricted cubic spline with knots at the 5th, 35th, 65th, and 95th percentiles of age is shown by equation image, in which c1(u) and c2(u) are cubic terms. The model was fit for a range of ages 66 to 94 years, but because of sparse data, ages 90 to 94 years were set to the value of the effect at age 90 years.

  • b

    Relative risk is for age 90 years relative to age 66 years.

  • c

    Relative risk is for year 2005 relative to 1995, for which the range is 1995 to 2005.

  • d

    Model was fit for a range of comorbidity scores of 0 to 5.125, but because of sparse data, scores of 1.5 to 5.125 were set to the value of the effect of a score of 1.5. The average contribution to the comorbidity score of 1 comorbidity was 0.43480.

  • e

    Relative risk is for a comorbidity score of 1.5 relative to a score of 0 (no comorbidity).

Age (restricted cubic spline)a3.5964b
 u-0.0121 (β1)0.03750.98790.7464
 C1(u)0.4363 (θ1)0.14081.54700.0020
 C2(u)-1.1623 (θ2)0.37770.31280.0021
Race    
 White (reference)1.0
 Black0.12690.09401.13530.1769
 Other-0.61970.16330.53810.0001
Marital status    
 Married (reference)1.0
 Single0.32550.06281.3847<0.0001
Year of diagnosis (linear)-0.08520.01180.4265c<0.0001
Comorbidity (linear)d0.0004<0.00011.0005e<0.0001
Stage by Gleason score    
 T1a, 2-7 (Reference)1.0
 T1a, 8-101.87510.24976.5213<0.0001
 T1b, 2-70.84900.15912.3373<0.0001
 T1b, 8-102.60850.145213.5792<0.0001
 T1c, 2-70.50460.12911.6563<0.0001
 1c, 8-101.46440.13304.3248<0.0001

Use of the output from the CSQS can best be illustrated by examples. In Table 3, a clinical scenario is given for each of 3 hypothetical prostate cancer patients, and inputs to the CSQS are given for each scenario along with outputs. The examples illustrate the importance of knowing the relative prognostic significance of both the cancer-specific and other-cause risks of dying, including life expectancy when patients are confronted with decisions regarding treatment of both their cancer and any existing comorbid conditions. Output from the CSQS for Example 2 is given in Figure 1. Curves are provided for the chance of dying of cancer, dying of other causes, and surviving. In addition to providing information based on the individual patient's profile, the second curve in each figure represents that for a person whose level of comorbidity is average for their age (ie, their chronologic age equals their HSAA).

thumbnail image

Figure 1. Sample output from the Cancer Survival Query System (CSQS) is shown for Example 2 in Table 3. Point Est indicates point estimate; 95% CI indicates 95% confidence interval.

Download figure to PowerPoint

Table 3. CSQS Query Examples for Prostate Cancer Along With Output From the System
 Example 1Example 2Example 3
  1. Abbreviations: 95% CI, 95% confidence interval; CSQS, Cancer Survival Query System.

Input
 Clinical scenarioA 67-year-old married white male is considering surgery vs active surveillance. He is concerned about balancing the need for cancer control against the side effects of surgery.A 74-year-old black widower has just completed external beam radiotherapy. He is interested in learning his prognosis as it relates to prostate cancer and his other comorbid conditions.A 66-year-old married white male has just undergone radical prostatectomy and was noted to have locally advanced disease with extension into the bladder neck. His physician tells him he is eligible for a clinical trial but he has numerous comorbid conditions and is trying to balance the benefits and risks of enrolling in the trial against his risk of dying of prostate cancer given his other illnesses.
 StagePretreatment clinical localized inapparent T1c N0 M0Pure clinical localized apparent T2 N0 M0Pathologic locally advanced pT4 pN0 pM0
 Gleason score688
 Age at diagnosis, y677466
 RaceWhiteBlackWhite
 Marital statusMarriedWidowerMarried
 Comorbid conditionsOld myocardial infarction, rhematologic diseaseDiabetes (without sequalae), old myocardial infarction, rheumatologic diseaseDiabetes (without sequalae), peripheral vascular disease
Output
 Health status-adjusted  age, y708177
Cancer-Specific Cumulative Crude Mortality Rate (%) (95% CI)
 3-y0.9 (0.7-1.2)4.1 (3.4-4.9)2.4 (1.1-4.6)
 5-y1.9 (1.5-2.3)7.1 (5.9-8.4)5.4 (2.5-10.0)
 10-y4.5 (3.5-5.7)12.6 (10.5-14.8)12.2 (5.4-21.7)
Other-Cause Cumulative Crude Mortality Rate (%) (95% CI)
 3-y8.0 (7.6-8.4)24.8 (23.7-25.9)15.6 (14.9-16.4)
 5-y14.3 (13.7-15.0)39.2 (37.7-40.8)26.7 (25.5-27.9)
 10-y33.3 (31.8-34.7)65.8 (63.3-68.1)53.0 (49.5-56.3)
Other-Cause Life Expectancy, Years
 1479

Validation

The discriminatory and calibration accuracy of the models was examined using time-dependent area under the receiver operating characteristic curve (AUC(t)) following DeLong et al24 and calibration plots suggested by Kattan et al.25 These 2 measures were computed via a 10-fold cross-validation. We followed the methods described by Lee et al26 to investigate how well the prostate cancer models used in the current study predicted crude cumulative mortality from cancer and from other causes.

Table 4 shows the AUC(t) for cancer (AUC1(t)) and from other causes (AUC2(t)) over 1, 3, 5, 7, and 10 years. For example, for patients with pretreatment clinical disease who were aged ≥ 66 years with comorbidity scores, the AUC(t) over 5 years was 0.85 for death from cancer and 0.72 for death from other causes, thereby indicating good discriminatory accuracy of the prostate cancer modeling. A value of 1 indicates perfect discrimination between those patients who survived and those who died over the 5-year period.

Table 4. Prostate Cancer Nomogram: Time-Dependent Area Under the ROC Curve for Death From Cancer and Death From Other Causes
 PretreatmentPure ClinicalPathological
YearAUC1(t)AUC2(t)AUC1(t)AUC2(t)AUC1(t)AUC2(t)
  1. Abbreviations: AUC1(t), area under the receiver operating characteristic curve for death from cancer; AUC2(t), area under the receiver operating characteristic curve for death from other causes; ROC, receiver operating characteristic curve.

Aged ≥66 Years With Comorbidity Scores
 1-y0.890.750.890.740.620.68
 3-y0.870.730.870.710.710.61
 5-y0.850.720.840.710.740.62
 7-y0.830.710.830.700.760.61
 10-y0.820.690.820.680.760.61
Aged <66 Years Without Comorbidity Scores
 1-y0.900.630.930.590.560.63
 3-y0.920.610.930.570.710.62
 5-y0.900.610.910.580.780.62
 7-y0.880.610.900.590.790.62
 10-y0.870.610.900.580.780.62

Figure 2 shows calibration plots for patients with pretreatment clinical prostate cancer with and without comorbidity scores. The crude cumulative mortalities from cancer and from other causes were computed over 1, 3, 5, 7, and 10 years. Patients were divided into 4 groups based on quartiles of the predicted crude cancer mortalities. For each quartile group, the mean of the predicted cancer mortalities was compared with a nonparametric estimate of crude mortality from cancer developed by Cronin and Feuer.4 In the same way, the mean of the predicted other-cause mortalities was compared with nonparametric estimates of crude mortality from other causes in each of the 4 quartile groups of the predicted crude mortalities. Pairs of the mean of the predicted crude mortality from cancer and the nonparametric estimate of crude mortality from cancer and the analogous pairs for other-cause mortality laid on or very close to the 45-degree angle line, which indicated that the prostate cancer prognostic models were good for predicting crude cumulative mortality from cancer and from other causes. Similar tables and plots for the other staging systems (ie, pure clinical and pathological) are provided in a technical report.14

thumbnail image

Figure 2. Calibration plots are shown for patients with pretreatment clinical prostate cancer with and without comorbidity scores. Each number and letter stands for the predicted time (1, 3, 5, 7, and 10 years) and quartile group (a, b, c, and d).

Download figure to PowerPoint

DISCUSSION

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. FUNDING SUPPORT
  8. REFERENCES

The CSQS presents survival information that is individualized as much as possible. A novel aspect of the independence assumption for competing risks is the ability to use different data sources for estimating other-cause mortality. The nomogram described herein is specifically designed to provide crude cause-specific survival estimates for prostate cancer patients aged 40 to 94 years, whereas the other-cause survival estimates for patients aged 66 to 94 years are a function of comorbid conditions. For example, localized prostate cancer is often a fairly slow-growing malignancy in which patients and providers are forced to weigh tumor characteristics (such as pathologic stage or grade) and host factors (such as physiological age, comorbid disease, and life expectancy) when choosing therapy. In fact, this need to consider the delicate balance of tumor aggressiveness (ie, the chance of dying of cancer) versus natural life expectancy (ie, the chance of dying of other causes) has lead clinicians to suggest that “active surveillance” (a therapeutic strategy in which frequent monitoring of the prostate-specific antigen level is coupled with repeat prostate biopsies to closely follow the cancer) may be a reasonable approach in lieu of more aggressive and morbid treatments such as radical prostatectomy or radiation therapy. Prior research has documented that patients place great value on controlling the potential spread of their cancer when choosing therapy for localized disease, but they also are almost equally concerned with treatment side effects that may influence their quality of life, such as impotence and incontinence. To this end, if the balance between a patient's natural life expectancy and tumor aggressiveness is such that prostate cancer is predicted to have a relatively minimal effect on overall survival, the issue of cancer control may be less important and patients may opt for less aggressive treatment in hopes of avoiding morbidity. The challenge, however, is to generate accurate information regarding tumor aggressiveness and natural life expectancy that is useful to providers and patients. In addition to being useful to cancer specialists, nomograms of this type can be used by internists and family physicians to make sure cancer patients pay proper attention to their other health conditions (which, surprisingly to some patients, may present a higher risk of death than their cancer). Although the inputs to the system require medical knowledge and judgment, the outputs were designed to be accessible to a wider audience (eg, pictograms).

Approximately 17% of patients aged < 66 years and 24% of patients aged ≥ 66 years were excluded from the analysis because of unknown prognostic variables or causes of death. Even if the remaining patients are not quite representative of SEER as a whole, all available prognostic variables were included in the regression analysis and therefore any bias would be minimized because it would have to occur conditionally based on the values of the regressor variables. For patients aged ≥ 66 years, there was also an issue regarding the extent to which the linked cases were representative of the total SEER patient experience. To examine this issue, the distributions of all prognostic variables for the linked versus nonlinked cases were compared and found to be very similar (analysis not shown).

Results from the CSQS for pathological staging should not be compared with those of patients who did not undergo surgery and thus were only staged clinically. Patients were not randomly assigned to a treatment, and even if one wanted to compare the 2 groups for the same values of all of the regressor variables in the models, the comparison would not be valid because stage and GS18 were determined primarily from the surgical findings in the pathological group and from the biopsy and clinical findings in the pure clinical group.

Although some usability testing was done in the development phase, extensive testing of the CSQS is being performed by Kaiser Permanente Colorado, one of the Centers for Excellence in Cancer Communication Research funded by the NCI.27 The objectives of this testing include investigating how members of health care teams view the general applicability, content and usability, and implementation potential of the tool; assessing the applicability of the tool across the cancer care spectrum (ie, types of providers and delivery systems); and assessing how patients and caregivers view the content and usability of the tool. Decisions regarding the mode of delivery of the CSQS will not be made until this testing is completed.

The patterns in the validation statistics presented in Table 4 indicate the importance of including comorbid conditions when assessing crude mortality from other causes. In all cases, the time-dependent AUC2(t) statistics were higher over the various follow-up periods for the patients with comorbidity data (aged ≥ 66 years) compared with those patients without comorbidity data (aged < 66 years), except for the pathological group. The reason for the somewhat lower statistics for the pathological group is likely related to the finding that to be eligible for surgery, patients generally must not have significant comorbid conditions. This homogeneity with respect to comorbid conditions makes it more difficult to accurately differentiate between patients with respect to their noncancer chance of death compared with the pretreatment and pure clinical groups (which are more heterogeneous). In general, the statistics for predicting the chances of cancer death are better than those for predicting other causes of death, and the statistics for patients aged < 66 years in the pretreatment and pure clinical groups have the highest AUC1(t), being on the order of 0.90 over the entire follow-up period.

A nomogram for colorectal cancer also has been developed, with details regarding the data analysis and validation provided in a technical report.14 Plans are also underway to develop nomograms for patients with breast and head and neck cancers.

FUNDING SUPPORT

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. FUNDING SUPPORT
  8. REFERENCES

Supported by the National Cancer Institute.

CONFLICT OF INTEREST DISCLOSURES

The authors made no disclosures.

REFERENCES

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. FUNDING SUPPORT
  8. REFERENCES
  • 1
    Surveillance, Epidemiology, and End Results (SEER) Program. http://seer.cancer.gov/. Accessed September 25, 2009.
  • 2
    Howlader N, Noone AM, Krapcho M, et al, eds. SEER Cancer Statistics Review, 1975-2008. Bethesda, MD: National Cancer Institute; 2011. http://seer.cancer.gov/csr/1975_2008/, based on November 2010 SEER data submission. Accessed July 5, 2011.
  • 3
    Ederer F, Axtell LM, Cutler SJ. The relative survival rate: a statistical methodology. Natl Cancer Inst Monogr. 1961; 6: 101-121.
  • 4
    Cronin KA, Feuer EJ. Cumulative cause-specific mortality for cancer patients in the presence of other causes: a crude analogue of relative survival. Stat Med. 2000; 19: 1729-1740.
  • 5
    Kattan MW, Scardino PT. Evidence for the usefulness of nomograms. Nat Clin Pract Urol. 2007; 4: 638-639.
  • 6
    Centers for Disease Control and Prevention. National Center for Health Statistics (US Life Tables). FastStats A to Z. http://www.cdc.gov/nchs/fastats/. Accessed September 25, 2009.
  • 7
    Centers for Medicare and Medicaid Services. http://www.cms.gov/. Accessed October 22, 2010.
  • 8
    Surveillance, Epidemiology, and End Results (SEER) Program. SEER Historical Coding and Staging Manuals. http://seer.cancer.gov/tools/codingmanuals/historical.html. Accessed September 27, 2009.
  • 9
    Surveillance, Epidemiology, and End Results (SEER) Program. SEER*Stat Database: Incidence-SEER 17 Regs, Nov. 2007 Sub (2000-2005) <Age Groups Including 85-89, 90-94, 95-99, and 100+, WITHOUT the Katrina/Rita Population Adjustment>-Linked To County Attributes-Total US, 1969-2005 Counties. Bethesda, MD: National Cancer Institute, Division of Cancer Control and Population Sciences, Surveillance Research Program, Cancer Statistics Branch; Year of publication 2009. Released February 2008, based on the November 2007 submission.
  • 10
    Surveillance, Epidemiology, and End Results (SEER) Program. SEER Registry Groupings for Analysis. http://seer.cancer.gov/registries/terms.html. Accessed April 13, 2011.
  • 11
    Howlader N, Ries LA, Mariotto AB, Reichman ME, Ruhl J, Cronin KA. Improved estimates of cancer-specific survival rates from population-based data. J Natl Cancer Inst. 2010; 102: 1584-1598.
  • 12
    Surveillance, Epidemiology, and End Results (SEER) Program. SEER Cancer-specific Death Classification. http://seer.cancer.gov/causespecific/index.html. Accessed October 31, 2010.
  • 13
    Prentice RL, Gloeckler LA. Regression analysis of grouped survival data with application to breast cancer data. Biometrics. 1978; 34: 57-67.
  • 14
    Feuer EJ, Lee M, Scoppa S, et al. Technical Report on the Cancer Survival Query System. Technical Report #2012-01. Bethesda, MD: Surveillance Research Program, National Cancer Institute; 2012. http://surveillance.cancer.gov/reports/. Accessed January 6, 2012.
  • 15
    Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis. 1987; 40: 373-383.
  • 16
    Klabunde CN, Potosky AL, Legler JM, Warren JL. Development of a comorbidity index using physician claims data. J Clin Epidemiol. 2000; 53: 1258-1267.
  • 17
    Klabunde CN, Legler JM, Warren JL, Baldwin LM, Schrag D. A refined comorbidity measurement algorithm for claims-based studies of breast, prostate, colorectal, and lung cancer patients. Ann Epidemiol. 2007; 17: 584-590.
  • 18
    Mariotto AB, Wesley MN, Cronin KA, Johnson KA, Feuer EJ. Estimates of long-term survival for newly diagnosed cancer patients: a projection approach. Cancer. 2006; 106: 2039-2050.
  • 19
    Percy C, Shanmugaratnam K, Whelan S, et al, eds. International Classification of Diseases for Oncology. 3rd ed. Geneva, Switzerland: World Health Organization; 2000.
  • 20
    Gleason DF. The Veteran's Administration Cooperative Urologic Research Group: histologic grading and clinical staging of prostatic carcinoma. In: Tannenbaum M, ed. Urologic Pathology: The Prostate. Philadelphia: Lea and Febiger; 1977: 171-198.
  • 21
    Carter HB, Allaf ME, Partin AW. Diagnosis and staging of prostate cancer. In: Wein AJ, Kavoussi LR, Novick AC, Partin AW, Peters CA, eds. Campbell-Walsh Urology. 9th ed. Philadelphia, PA: Saunders; 2007: 2912-2931.
  • 22
    Greene FL, Page DL, Fleming ID, et al, eds. AJCC Cancer Staging Manual. 6th ed. New York, NY: Springer-Verlag; 2002.
  • 23
    Durrleman S, Simon R. Flexible regression models with cubic splines. Stat Med. 1989; 8: 551-561.
  • 24
    DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988; 44: 837-845.
  • 25
    Kattan MW, Heller G, Brennan MF. A competing-risks nomogram for sarcoma-specific death following local recurrence. Stat Med. 2003; 22: 3515-3525.
  • 26
    Lee M, Cronin KA, Gail MH, Feuer EJ. Predicting the absolute risk of dying from colorectal cancer and from other causes using population-based cancer registry data. Stat Med. 2012; 31: 489-500.
  • 27
    National Cancer Institute, Division of Cancer Control and Population Sciences. Centers of Excellence in Communications Research Initiative. http://cancercontrol.cancer.gov/hcirb/ceccr/ceccr-index. html. Accessed July 6, 2011.