|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
© 2001 American Society for Clinical Oncology A Simple Stratification Factor Prognostic for Survival in Advanced Cancer: The Good/Bad/Uncertain IndexByFrom the Mayo Clinic and Mayo Foundation, Rochester, MN; Altru Health Systems, Grand Forks, ND; Duluth Community Clinical Oncology Program, Duluth, MN; Allan Blair Cancer Center, Regina Saskatchewan, Canada; Ochsner Community Clinical Oncology Program, New Orleans, LA; Cedar Rapids Oncology Project Community Clinical Oncology Program, Cedar Rapids, IA; Sioux Community Cancer Consortium, Sioux Falls, SD; Illinois Oncology Research Association Community Clinical Oncology Program, Peoria, and Carle Cancer Center Community Clinical Oncology Program, Urbana, IL; and Toledo Community Hospital Oncology Program Community Clinical Oncology Program Toledo, OH. Address reprint requests to Jeff A. Sloan, PhD, Mayo Clinic, 200 1st St, SW, Rochester, MN 55905; email: jsloan{at}mayo.edu
PURPOSE: This article summarizes the third step of a research program to identify variables that supplement the predictive power of the the Eastern Cooperative Oncology Group (ECOG) performance status (PS) for survival. The objective was to produce a simple, practical, stratification factor for phase III oncology clinical trials involving patients with advanced malignant disease. PATIENTS AND METHODS: A questionnaire was administered to 729 patients with metastatic colorectal or lung cancers. Patients provided a Karnofsky index and appetite rating while physicians provided a survival estimate and the ECOG-PS. Scores for each item were categorized as having a positive, neutral, or negative indication for survival. A patient was classified as having a relatively good prognosis if three or more of the four items showed a positive indication, a bad prognosis if three or more items were negative, and an uncertain prognosis otherwise (Good/Bad/Uncertain [GBU] index). RESULTS: The GBU index improved on the prognostic power of a Cox model quartile index and PS alone and increased the accuracy of survival classification estimates by 5% to 10% more than ECOG-PS alone. For patients with PS of 0 or 1, significant survival patterns exist between GBU groups (P= .002 and .0001, respectively). CONCLUSION: The GBU index may be recommended as a supplementary stratification factor for certain future phase III trials in metastatic lung or colorectal cancer where patient heterogeneity is a particular concern. The GBU represents a relatively modest increase to the cost and patient burden of a clinical trial given the additional control that is achieved over the potentially confounding concomitant to the treatment variable.
STRATIFICATION FACTORS are useful components of clinical trials because they allow for the simple interpretation of the results for primary end points in the presence of potentially confounding concomitant variables.1-10 Because it is not practical to stratify all possible confounding influences,6,7,9,10 the challenge is to identify a minimal clinically relevant subset of variables to use as stratification factors.8,9 In cancer populations, that goal is to identify a set of clinical indicators that collectively will reflect the extent of disease and degree of suffering a patient is experiencing. In addition, patients may be classified according to their relative likelihood of doing well in a trial. The Eastern Cooperative Oncology Group (ECOG) performance status (PS) is one of the most useful stratification variables in oncology clinical trials.11-14 It is easy to complete and takes a modest effort to produce a reasonably valid and reliable score reflecting the degree of physical mobility evidenced for a particular patient.12,14 The ECOG-PS repeatedly has been shown to be prognostic for survival in cancer patients.15 However, physical function cannot by itself accurately identify which patients will do well.14-16 Other related instrumentation targeted specifically to patient quality of life has also demonstrated prognostic capability.16,17 Much has been written about the value of patient opinion on prognosis and quality of life.18-23 Clinician opinion regarding patient outcome has also been seen to be prognostic for survival.24 One would intuitively surmise that it would be useful to supplement the established prognostic ability of the ECOG-PS score by a small number of additional items to provide a more equitable balance across treatment arms in regard to the proportion of patients who will do well or do poorly. The challenge is to keep the subset of items to a minimum while supplementing the prognostic power of the ECOG-PS. Using models created in steps 1 and 2, we present the results of model validation using patient data from step 3. This report outlines the construction of a simple prognostic index for survival. Our results indicated that the Good/Bad/Uncertain (GBU) index shows promise as a stratification factor for future oncology clinical trials.
This research involved three steps. The first step was the development of a 30-item questionnaire that was pilot-tested on a sample of 186 cancer patients with advanced, incurable, malignant disease. Both patients and physicians completed the questionnaire for purposes of refinement. The second step involved variable selection and model building using data from 1,560 patients. This step used a revised questionnaire with a sample of patients who had been entered onto North Central Cancer Treatment Group (NCCTG) clinical trials. The previously published results indicated that the following subset of four items was prognostic for survival: physician assessment of (1) performance status and (2) likelihood of survival, (3) patient assessments of performance status, and (4) patients’ report of appetite.25 Finally, the third step established index validation that constituted the goal of the present report. The items selected for step 3 in this report were the physician-completed ECOG-PS, physician estimate of survival, and the patient-completed Karnofsky score and appetite assessment (Table 1). Each of the four items was measured on an ordinal scale.26
To construct the GBU index, we started by using the Cox proportional hazards model from step 2 of the research process as the basis for variable selection.25 With the step 2 sample, regression coefficients were derived for each of the response categories of the four prognostic variables and were later used for generating Cox scores for each patient. The quartiles of these Cox scores formed our original classification groupings. The Cox scores were obtained by multiplying the regression weights by the values observed for each prognostic variable and adding the resulting products. The final score was the natural base power of the resulting sum. We used these Cox scores from previous steps to validate the GBU index. The GBU index arises out of a practical consideration that the Cox scores are difficult to implement clinically. The weights were indicative of each component item being an equally important prognostic indicator. Hence, we attempted to construct a simpler and clinically more viable index by classifying results on each indicator as positive, negative, or neutral in implication. The validation step after this work included a sensitivity analysis where we assessed the number of items needed for a reliable prognostic factor. We established three rules for selecting the prognostic variables named "two of four," "three of four" and "four of four." For example, the rule "three of four" signifies that a patient has at least three of the four prognostic variables, as previously described, with positive or negative implications. In this scenario, a patient was classified as having a relatively good prognosis if three or more of the four items showed a positive indication, a relatively bad prognosis if three or more items were negative, and an uncertain prognosis otherwise. This classification system was referred to as the GBU index. The next validation step was to fit survival curves to assess the prognostic power of each rule. The validation step also included (1) considering survival curves by GBU levels separately for patients with ECOG-PS categories of 0, 1, and 2/3; (2) looking at categories of actual survival rates by GBU level; and (3) reanalyzing protocols, including the GBU index. Standard approaches to psychometric validation were used.
The sample used in step 3 involved 729 patients who were on NCCTG clinical trials. Table 2 presents the distribution of baseline factors. Compared with previous steps, this report included fewer female, black, and advanced-disease cancer patients. The diversity of the sample in this step was actually an advantage because it provided further generalizability beyond the characteristics of the initial samples in previous steps. Using this sample, the model-building step included the use of Cox proportional hazards models.
Using results from step 2, we found that subpopulations formed using the quartiles of the Cox scores produced statistically significantly different survival curves. Figure 1 presents the survival curves for the step-3 data set of 729 patients. This figure indicates that the Cox model quartiles constructed on the step-2 data are validated for step-3 data by the clear separation among the four quartiles’ survival curves.
In order to assess the contribution of the four items to patient survival, we calculated their Cox model weights. A Cox proportional hazards model was created. Table 3 presents the standardized weights for each of the four variables. The standardized coefficients allowed a direct comparison of the coefficients, regardless of the different units of measurement for each item. Two results were immediately evident. First, the weights were all roughly equivalent to one another. This would indicate that using the Cox weights produces relative rankings of patient survival that were basically the same as if the four items were weighted equally. Second, the calculation of Cox scores for individual patients required a substantial amount of calculation beyond what is practical in the clinical environment or at the randomization of a patient onto a clinical trial. These two facts combined suggested the possibility of constructing a simple index that would be clinically practical and yet retain the prognostic power of the more complicated Cox scores. The Cox weights for each of the four variables in Table 3 were basically identical, which indicated that there was no particular dimension among the four items containing any more prognostic power than the others. Hence, we decided to examine the implications for survival for each score.
The schema for comparing the four items to generate the GBU index is included in Table 4. Table 4 contains the mapping of the individual scores for each item onto a classification system based on whether each score had a positive, negative, or neutral implication for a patient’s survival. The percentage of patients who had positive implications for all four variables was 11% (162 of 1,538 patients). Similarly, only 5% (80) of the patients had all four variables scored as negative implications. Hence, neither absolute was useful as a basis to predict patient survival. We assessed the efficiency of this strategy by plotting a survival curve for each of the GBU categories.
Figure 2 plots the survival curves for the three GBU categories. We observed a separation of the survival curves, indicating that the GBU index had substantive prognostic power for patient survival. In addition, we compared the percentage of patients allocated to good, bad, and uncertain categories. The percentages comparing steps 2 and 3 were remarkably similar (data not shown), leading to a conclusion similar to that shown in Fig 2. The three groups were clearly identifiable and the GBU index had a good prognostic power. For example, for patients with a survival time of less than 1 year, 78% of step 2 and 76% of step 3 patients were allocated to the bad category.
In the validation process, we assessed the number of items that were necessary to provide reasonable prognostic power. The outstanding issues were to validate the index and to determine whether the GBU index provided more prognostic power than the more complex Cox scores approach and more than simply using the ECOG-PS. For this reason, we performed a sensitivity analysis for the GBU index using a "two of four" rather than the "three of four" rule. The number of patients that would be classified as good under this characterization becomes impractical, since 56% of the patients are allocated to the good category. Similarly, using a "four of four" rule produced an overabundance of uncertain classifications (data not shown). The complex approach uses the Cox quartiles and divides the population into four categories as presented in Table 5. The GBU index provided a simpler way to allocate each individual into three categories, as shown in Table 6. The comparison of the GBU index to more complex scores with four variables can be observed by comparing the percentages from Tables 5 and 6. For example, using stage 2 data for patients with survival times of 1 to 2 years, 38% of patients were in the good category in Table 6. This percentage is comparable to 35% from the first quartile in Table 5. As observed using the GBU index, we obtained similar information to that of the more complex Cox quartiles approach in less time.
A survival curve similar to Fig 2 was plotted to observe the prognostic power of the Cox quartiles approach. Figure 3 presents the survival curves broken down by the various ECOG-PS categories. The comparison of the separation of the curves for the various ECOG-PS categories with the GBU curves in Fig 2 indicated that the GBU index produced curves with a greater degree of separation. Figure 3 shows a good separation for categories 0, 1, and 2 to 1,000 days. In addition, we observed by collapsing categories 2 to 3, similar to Fig 3, that separation of the survival curves did not improve (not shown).
We took a further validation step to compare the GBU index with different ECOG-PS categories to assess its prognostic ability. Figure 4 illustrates that the GBU index provided additional information to that of ECOG-PS. This figure shows different PS levels and then compares survival curves among GBU categories within each PS category. Figure 4 provides further evidence by demonstrating that survival curves for ECOG-PS individual categories (0, 1, and 2-3) can be separated further into significantly different subpopulations based on the GBU index classification. Figure 4A indicates significant survival differentiation for patients with PS = 0 between good versus uncertain patients (P = .002). Figure 4B similarly shows that, among patients with PS = 1, the GBU index identified good, bad, and uncertain prognosis subpopulations, which have significantly different survival patterns (P = .0001). Figure 4C illustrates that the GBU index did not significantly differentiate survival curves for patients with ECOG-PS of 2 to 3.
A series of additional Cox models were obtained to assess the ability of the GBU index compared with ECOG-PS. For example, a Cox proportional hazards model examined the relative prognostic power of the GBU index to the ECOG-PS. Estimates for the Cox models for the GBU index compared with ECOG-PS indicated that the PS alone (relative risk [RR] = 1.38; P = .0001) and GBU alone (RR = 1.59; P = .0001) had more predictive power when combined into the model simultaneously (PS with RR = 1.15, P = .03; GBU with RR = 1.43, P = .0001) with a significant  2 P value (P < .0001). These results indicate that the model with ECOG-PS alone provides substantial predictive power for survival when entered into the model without the GBU index, as does the GBU index when entered into the model without the ECOG-PS. Most importantly, when both variables are entered into the model, the GBU provides more supplementary significant predictive power for survival than what is achieved by the ECOG-PS alone. An additional test was conducted for the GBU index by reanalyzing individual clinical trials from among the stage III data set of 729 patients. Two trials provided sufficient numbers of patient data to examine whether the GBU index would have changed the ultimate outcome of the trials. We obtained Cox proportional hazards models for the NCCTG clinical trials 89-24-51 and 89-46-52 (data not shown). It is comforting to note that the overall conclusions of the trials would not have changed had the GBU index been used as a stratification factor. However, the significance of the treatment effect in trial 89-24-51 became substantially more nonsignificant once the results were analyzed as conditional on the GBU index. Furthermore, the GBU index added significantly to the prediction of survival, even in the presence of the ECOG-PS in the models.
The results shown constitute relevant validation steps of the prognostic ability of the GBU index. Collectively, the four items have been tested on more than 2,000 patients with advanced cancer. Not only was the initial result of the usefulness of the four items confirmed, but a simple, practical means of operationally incorporating this information was constructed. A similar application has been used in constructing the International Prognostic Index for Non-Hodgkin’s Lymphoma.27,28 This index can be used to predict the likelihood of relapses as well. It is important to emphasize that the GBU index is intended as a supplement to the ECOG and/or Karnofsky performance status indices, not a replacement. Our results indicate that the ECOG-PS is a valuable and strongly prognostic factor for survival among advanced cancer patients. Results also show, however, that even with the ECOG-PS used to stratify enrollment onto a clinical trial, the GBU index provides further prognostic power for survival. This result suggests that the additional confounding variables that are removed by the GBU index may well alter the statistical significance in some trials. Having said this, it is also clear that large randomized trials involving hundreds or thousands of patients can rely on randomization to balance all the prognostic factors among patient groups. The gain in stratification and variance reduction must be balanced against pragmatic, cost, and patient burden considerations. At present, the ECOG-PS is collected routinely in many oncology clinical trials. The addition of the three extra items needed to complete the GBU index does require some extra effort. The patient has to respond to two additional questions, and the physician must also complete two additional items before a patient is randomized onto a trial. We have implemented the GBU index as a stratification factor in a number of NCCTG oncology clinical trials. To date, no difficulties have been reported in obtaining the GBU score, but some clinicians have commented on the extra effort that it takes. We plan to explore alternative logistics for administering the GBU in order to minimize the additional effort. The GBU hopefully is a reasonable and acceptable addition to the running of an oncologic clinical trial, taking into account that we can achieve a further reduction in potential concomitant confounding elements with the treatment variable better than that achieved by the ECOG-PS. The results showed that including further questions to complete the GBU index would reduce confounding influences. In the decision to include any stratification factors, even the ECOG-PS must be taken with due consideration to balance the additional power against the added patient and resource burdens. Several review articles have discussed the need for stratification in clinical trials.26,29-32 The general consensus is that randomization often removes the concern for systemic imbalance among confounding covariates. The GBU index ensures control of four potentially confounding influences. In trials where such influences are unlikely to be profound, the GBU would represent an unnecessary addition. The GBU index is likely most suitable in multisite clinical trials in which patient heterogeneity is a considerable challenge.
Additional participating physicians and institutions include: Suresh Nair, MD, Geinsinger Clinic and Medical Center Community Clinical Oncology Program, Danville, PA; Roscoe F. Morton, MD, Iowa Oncology Research Association Community Clinical Oncology Program, Des Moines, IA; Ferdinand Addo, MD, Quain and Ramstad Clinic, Bismarck, ND; Harold E. Windschitl, MD, Centra Care Clinic, St Cloud, MN; Larry P. Ebbert, MD, Rapid City Regional Oncology Group, Rapid City, SD; John C. Michalak, MD, Siouxland Hematology-Oncology Associates, Sioux City, IA; Ralph Levitt, MD, Meritcare Hospital Community Clinical Oncology Program, Fargo, ND; and James A. Mailliard, MD, Nebraska Oncology Group, Creighton University of Nebraska Medical Center and Associates, Omaha, NE.
Supported in part by Public Health Service grant nos. CA-25224, CA-37404, CA-15083, CA-35269, CA-35113, CA-35272, CA-52352, CA-35103, CA-37417, CA-63849, CA-35448, CA-35101, CA-35195, CA-35415, and CA-35103.
This study was conducted as a collaborative trial of the North Central Cancer Treatment Group and Mayo Clinic.
1. Pocock SJ: Allocation of patients to treatment in clinical trials. Biometrics 35: 183-197, 1979[Medline] 2. Begg CB, Iglewicz B: A treatment allocation procedure for sequential clinical trials. Biometrics 36: 81-90, 1980[Medline] 3. Simon R: Patient heterogeneity in clinical trials. Cancer Treatm Rep 64: 405-410, 1980 4. Brown BW Jr: Designing for cancer trials: Selection of prognostic factors. Cancer Treatm Rep 64: 499-502, 1980 5. Meier P: Stratification in the design of a clinical trial. Control Clin Trials 1: 355-361, 1981[Medline] 6. Kalish LA, Begg CB: Treatment allocation methods in clinical trials: A review. Stat Med 4: 129-144, 1985[Medline] 7. Brandmaier RM, Aydemir U, Ansari H, et al: Do we always need clinical stratification in multicenter clinical trials? Results of a simulation study. Proc SAS Users Group Int Conf 17: 1439-1444, 1992 8. Therneau TM: How many stratification factors are too many to use in a randomization plan? Control Clin Trials 14: 98-108, 1993[Medline] 9. Feinstein AR, Landis JR: The role of prognostic stratification in preventing the bias permitted by random allocation of treatment. J Chronic Dis 29: 277-284, 1976[Medline]
10.
Lininger LG, Green MH, Byar SB, et al: Comparison of four tests for equality of survival curves in the presence of stratification and censoring. Biometrika 66: 419-428, 1979 11. Karnofsky DA, Burchenal JH: The clinical evaluation of chemotherapeutic agents in cancer, in Macleod CM (ed): Evaluation of Chemotherapeutic Agents. New York, NY, Columbia University, 1949, pp 199-205 12. Schag CC, Heinrich RL, Ganz PA: Karnofsky performance status revisited: Reliability, validity, and guidelines. J Clin Oncol 2: 187-193, 1984[Abstract] 13. Hutchinson TA, Boyd NF, Feinstein AR, et al: Scientific problems in clinical scales as demonstrated in the Karnofsky index of performance status. Cancer 45: 2220-2224, 1980[Medline] 14. Mor V, Laliberte L, Morris JN, et al: The Karnofsky performance status scale: An examination of its reliability and validity in a research seting. Cancer 53: 2002-2007, 1984[Medline] 15. Coates A, Gebski V, Signorini D, et al: Prognostic values of quality of life scores during chemotherapy for advanced breast cancer. J Clin Oncol 10: 1833-1838, 1992[Abstract] 16. Weeks J: Quality of life assessment: Performance status upstaged? J Clin Oncol 10: 1827-1829, 1992 (editorial)[Medline] 17. Degner LF, Sloan JA: Symptom distress in newly diagnosed ambulatory cancer patients and as a predictor of survival in lung cancer. J Pain Symptom Manage 10: 423-431, 1995[Medline] 18. Cassileth BR, Lusk EJ, Miller DS, et al: Psychosocial correlates of survival in malignant disease. N Engl J Med 312: 1551-1555, 1985[Abstract] 19. Cella DF, Bonomi AE: Measuring quality of life: 1995 update. Oncology 9: 47-60, 1995 (suppl)[Medline]
20.
Moinpour CM, Fiegl P, Metch B, et al: Quality of life end points in cancer clinical trials: Review and recommendations. J Natl Cancer Inst 81: 485-495, 1989 21. Patrick D, Deyo R: Generic and disease-specific measures in assessing health status and quality of life. Med Care 27: S217-S232, 1989 (suppl 3)[Medline] 22. Spitzer WO: State of science 1986: Quality of life and functional status as target variables for research. J Chronic Dis 40: 465-471, 1987[Medline] 23. Tamburini M, Brunelli C, Rosso S, et al: Prognostic value of quality of life scores in terminal cancer patients. J Pain Symptom Manage 11: 32-41, 1996[Medline] 24. Slevin ML, Plant H, Lynch D, et al: Who should measure quality of life: The doctor or the patient? Br J Cancer 57: 109-112, 1988[Medline] 25. Loprinzi CL, Laurie A, Wieand HS, et al: Prospective evaluation of prognostic variables from patient-completed questionnaires. J Clin Oncol 12: 601-607, 1994[Abstract] 26. Kalish LA, Begg CB: Treatment allocation methods in clinical trials: A review. Stat Med 4: 129-144, 1985
27.
A predictive model for aggressive non-Hodgkin’s lymphoma: The International Non-Hodgkin’s Lymphoma Prognostic Factors Project. N Engl L Med 329: 987-994, 1993 28. Hasenclever D, Diehl V: A prognostic score for advanced Hodgkin’s disease: International Prognostic Factors Project on Advanced Hodgkin’s disease. N Engl J Med 21: 1506-1514, 1998 29. Brown BW Jr: Designing for cancer trials: Selection of prognostic factors. Cancer Treat Rep 64: 499-502, 1980[Medline] 30. Feinstein AR, Landis JR: The role of prognostic stratification in preventing the bias permitted by random allocation of treatment. J Chronic Dis 29: 277-284, 1976 31. Meier P: Stratification in the design of a clinical trial. Control Clin Trials 1: 355-361, 1981 32. Longford NT: Selection bias and treatment heterogeneity in clinical trials. Stat Med 18: 1467-1474, 1999[Medline] Submitted September 16, 1999; accepted May 1, 2001.
This article has been cited by other articles:
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2001 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
TOP
ABSTRACT
INTRODUCTION
