The ten-year rule revisited: accuracy of clinicians’ estimates of life expectancy in patients with localized prostate cancer

ADULT UROLOGY

THE TEN-YEAR RULE REVISITED: ACCURACY OF CLINICIANS’ ESTIMATES OF LIFE EXPECTANCY IN PATIENTS WITH LOCALIZED PROSTATE CANCER MURRAY D. KRAHN, KAREN E. BREMNER, JAMIL ASARIA, SHABBIR M. H. ALIBHAI, ROBERT NAM, GEORGE TOMLINSON, MICHAEL A. S. JEWETT, PADRAIG WARDE, AND GARY NAGLIE

ABSTRACT Objectives. To determine the accuracy of clinicians’ predictions of life expectancy in patients with localized prostate cancer, when provided with information about age and comorbidity, and to determine whether life expectancy estimates predict the choice of initial treatment. Methods. A survey was sent by facsimile to 191 Canadian urologists and radiation oncologists asking them to estimate the life expectancy and choose the initial therapy (radical prostatectomy, radiation, or watchful waiting) for 18 patient scenarios: two prostate cancer scenarios, each with three ages and three levels of comorbidity. Results. Life expectancy estimates were accurate within 1 year of the projections of a Markov model for 31% of the clinicians’ responses and accurate within 3 years for 67% of the responses. The average prediction error ranged from 2.4 to 5.2 years. The life expectancy was correctly estimated as being greater than or less than 10 years in 82% of responses. Ten years was the minimal life expectancy for recommending surgery and within the range (5 to 15 years) in which radiation was recommended. Conclusions. Clinicians can use age and comorbidity to predict the life expectancy of patients with localized prostate cancer with a modest degree of overall accuracy, but with sufficient accuracy to use the “10-year rule.” Life expectancy estimates are strongly associated with treatment choice. The appropriateness of the 10-year rule remains to be determined. UROLOGY 60: 258–263, 2002. Crown copyright © 2002. Published by Elsevier Science Inc.

T

he “10-year rule” is a frequently cited heuristic in prostate cancer clinical decision making. One version states that curative therapy (surgery or radiation) for localized prostate cancer should only be offered to patients with a life expectancy of at least 10 years.1–3 Advocates of this rule assume that life expectancy can be predicted accurately, but little evidence supports this belief. One study From the Departments of Medicine, Surgery (Urology), Health Policy, Management, and Evaluation, Public Health Sciences, Radiation Oncology and School of Medicine, University of Toronto; Divisions of General Internal Medicine and Clinical Epidemiology and Urology, Toronto General Hospital; and Department of Radiation Oncology, Princess Margaret Hospital, University Health Network, Toronto, Ontario, Canada Reprint requests: Murray D. Krahn, M.D., M.Sc., Department of Medicine, Toronto General Hospital, 200 Elizabeth Street, Room ENG-248, Toronto, Ontario M5G 2C4, Canada Submitted: August 23, 2001, accepted (with revisions): March 14, 2002

258

demonstrated that 80% of radical prostatectomy patients at a single center actually had a life expectancy greater than 10 years.4 Other studies in patients terminally ill with cancer have shown clinicians’ life expectancy estimates to be only 20% to 49% accurate.5– 8 No studies have systematically examined the accuracy of clinicians’ survival predictions in a broad range of patients with prostate cancer. We therefore sought to answer two questions. First, how accurate are clinicians in predicting the life expectancy of patients with localized prostate cancer? Second, do predictions of life expectancy correspond with treatment choices? MATERIAL AND METHODS All members of the College of Physicians and Surgeons of Ontario actively practicing adult urology as a specialty (www.cpso.on.ca) and all radiation oncologists in Canada currently treating genitourinary cancers according to the On-

Crown copyright © 2002. Published by Elsevier Science Inc. ALL RIGHTS RESERVED

PII S0090-4295(02)01712-0

tario Cancer Institute (Toronto) were surveyed by facsimile. Contact information and year and place of graduation were obtained from the Canadian Medical Directory.9 The survey questionnaire described two prostate cancer scenarios: Stage T1c with a Gleason score of 6 and a prostatespecific antigen level of 5.6 ng/mL, and Stage T2a with a Gleason score of 7 and a prostate-specific antigen level of 10.2 ng/mL. These scenarios represent moderately aggressive clinically localized tumors, for which the optimal treatment was thought to be uncertain. Within each prostate cancer scenario, age and comorbidity were modified. Patients of three different ages (55, 65, and 75 years) were each described as having (a) no comorbid illness or functional impairment (Index of Coexistent Disease [ICED]10 ⫽ 0); (b) myocardial infarction 5 years previously, angina when climbing one flight of stairs, and been prescribed aspirin, beta-blockers, and nitrates by a cardiologist (ICED ⫽ 2); (c) symptomatic chronic obstructive pulmonary disease with shortness of breath when brushing teeth or getting dressed, recurrent lung infections, myocardial infarction 5 years previously, angina when climbing one flight of stairs, and been prescribed aspirin, calcium channel blockers, nitrates, and ipratropium bromide by an internist (ICED ⫽ 3). Clinicians were asked to select one of three initial management options (watchful waiting, any type of radiotherapy, or radical prostatectomy, each with or without hormonal therapy) for each of the nine patient descriptions within each prostate cancer scenario and to estimate the life expectancy of the patient, disregarding the patient’s prostate cancer. Clinicians were requested to return the completed questionnaire, or a reason for not completing it, by facsimile. A second facsimile was sent if no reply was received within 2 weeks. Nonrespondents were then mailed the questionnaire and a stamped, addressed return envelope. Because approximately twice as many urologists were surveyed as radiation oncologists, we were able to ease their response burden by randomly sending most of the 212 urologists one prostate cancer scenario with 9 patient descriptions; the 98 radiation oncologists were each sent all 18 patient descriptions.

MODEL OF LIFE EXPECTANCY The true life expectancies of the patients described in the scenarios were derived from a decision-analytic model of life expectancy developed by our group.11,12 A two-state (living, dead) Markov model was created to model life expectancy using two variables—age and comorbidity. Cohorts of patients were assigned an initial age corresponding to the ages used in our scenarios (55, 65, and 75 years). The annual agestratified mortality rate for Canadian men was obtained from life tables,13 and the increased risk of mortality owing to comorbidity was modeled as a multiplier of it. The annual hazard ratio of dying and 95% confidence interval for each different ICED level, relative to having no comorbidity, was derived from a cohort of conservatively treated men with prostate cancer.14 The hazard ratio represented the increased risk of dying at a given age for a given level of comorbidity. Monte Carlo simulation was used to derive 95% confidence intervals around each life expectancy estimate using the lognormal distribution for the comorbidity covariate and 1000 cycles of 50 trials.

STATISTICAL ANALYSIS Demographic differences between urologists and radiation oncologists were determined using chi-square and analysis of variance tests for categorical and continuous variables, respectively, using Statistical Package for the Social Sciences for Windows.15 To determine the factors predicting the accuracy of the cliUROLOGY 60 (2), 2002

nicians’ estimates of life expectancy, we used linear regression analysis with the difference between the life expectancy predicted by the Markov model (the “true” life expectancy) and the clinicians’ estimates as the dependent variable, using SPlus 2000 Professional software.16 To account for the uncertainty in the model-based true life expectancy, we performed the regression analysis for each of 1000 sets of true values generated from the model (one set ⫽ one life expectancy for each scenario). Also, to summarize the clinicians’ errors in predicting life expectancy, we computed the root mean squared error of their life expectancy estimates for each scenario. This is the square root of the average of the following: (clinician estimated life expectancy ⫺ true life expectancy)2 across clinicians. The root mean squared error is the average difference between the true (model’s) life expectancy value and a typical clinician’s life expectancy estimate. To determine the factors predicting the treatment recommendation (radical prostatectomy versus others), we performed logistic regression analysis with the WinBUGS package.17 For both analyses, we used a mixed model with random effects for clinicians. The clinician’s status as a urologist or radiation oncologist was modeled as a between-subjects variable and age, tumor stage, comorbidity, and life expectancy were modeled as within-subject variables. This recognizes that each clinician may have a different “baseline” response (ie, probability of recommending a treatment or error in estimating life expectancy) that is modified by the characteristics of the patient scenario under consideration.

RESULTS RESPONSE Of the 310 clinicians surveyed, 14 replied that they did not see patients with prostate cancer, 6 had retired, and 13 could not be located. Completed surveys were returned by 69% of the eligible 277 clinicians (138 urologists [73%], 53 radiation oncologists [60%], P ⬍0.05). Most of the urologists (n ⫽ 115) completed only 9 patient descriptions, and 23 urologists and all but 4 radiation oncologists completed all 18 patient descriptions. Most of the respondents were male; more radiation oncologists (15.1%) than urologists (2.2%) were women (P ⬍0.01). More urologists had graduated from a Canadian university (70% versus 50%, P ⫽ 0.01). LIFE EXPECTANCY ESTIMATES Clinicians’ life expectancy estimates fell as patient age and comorbidity increased, similar to the pattern of the Markov model (Fig. 1). For all scenarios, 31% of the clinicians’ estimates were within ⫾1 year of the mean of the model’s value for that scenario, 48% of their estimates were within 2 years of the mean of the model’s values, and 67% of their estimates were within 3 years of the mean of the model’s values. Fifty-five percent of clinicians’ life expectancy estimates for all scenarios were within the 95% confidence interval of the model’s estimates (excluding ICED ⫽ 0 scenarios, for which life expectancies were certain as they were from the population). The average error for any clinician’s estimate (root mean squared error) 259

FIGURE 1. Distributions (percentage of total) of clinicians’ estimates of life expectancies of patients according to comorbidity and age (represented by columns) compared with median and 95% confidence interval estimates of a model12 (represented by box plots [median, 25th, and 75th percentile values] with whiskers [2.5th and 97.5th percentiles]). No distributions shown for median of model’s estimates for healthy patients because these data were derived from the actual population. Root mean squared error (RMSE) represents the mean error for clinicians compared with the reference standard estimate (median of model predictions). MI ⫽ myocardial infarction; COPD ⫽ chronic obstructive pulmonary disease; LE ⫽ life expectancy.

ranged from 2.4 to 5.2 years in the various scenarios. However, clinicians were surprisingly accurate in estimating patients’ life expectancy around the 10-year benchmark. Overall, 82% of the clinicians’ estimates correctly classified patients as having a life expectancy less than or greater than 10 years. Even for the two scenarios with predicted life expectancies close to 10 years, most clinicians (86% for the healthy 75 year old, and 62% for the 65 year old with ICED ⫽ 2) correctly classified the life expectancy as less than 10 years (model’s predicted life expectancy ⫽ 9.0 and 8.9 years, respectively; Fig. 1). The amount of error in the clinicians’ estimates was not predicted by age, tumor stage, comorbidity, or clinician specialty (Table I). The only significant factor determining error was the model’s life expectancy; the clinicians’ error increased by 1.08 years per year of the model’s prediction. TREATMENT DECISIONS Age, comorbidity, and estimated life expectancy, but not tumor stage, were significant independent predictors of treatment decision (Table I). For ex260

ample, a 55-year-old patient was three times as likely to be offered surgery as a 65-year-old patient and more than 300 times likely than a 75-year-old man. Figure 2 shows the clinicians’ recommendations for radical prostatectomy according to their estimates of patient life expectancy. Ten years appeared to be a critical threshold, as surgery was almost never recommended when the life expectancy was shorter. Patients with life expectancies between 5 and 15 years were most likely to be offered radiotherapy (Fig. 3). Watchful waiting was more often recommended when the estimated life expectancy was less than 10 years and was recommended by more than 80% of the clinicians when it was 3 years or less. Urologists were 27 times more likely to recommend surgery than were the radiation oncologists (Table I). The odds ratios varied with the patient scenario. Urologists were 75 times more likely than radiation oncologists to recommend surgery for the healthy 55 year old, but only 4 times as likely to do so for the 55 year old with an ICED of 2. UROLOGY 60 (2), 2002

TABLE I. Predictors of error in life expectancy estimates (deviation from model) and choice of radical prostatectomy as a treatment option Predictors of Error in LE Estimates* Years

Factor Age (yr) 55 65 75 Tumor stage and grade T1c, Gleason score 6 T2a, Gleason score 7 Comorbidity ICED ⫽ 0 ICED ⫽ 2 ICED ⫽ 3 Model’s LE estimates (per year of life) Clinician LE estimates (per year of life) Clinician specialty Radiation oncologist Urologist

95% CI

Odds of Recommending Radical Prostatectomy Odds Ratio

95% CI

0 1.285 1.262

— ⫺1.602–6.48 ⫺2.831–9.043

1 0.305 0.003

— 0.161–0.567 0.001–0.012

0 ⫺0.170

— ⫺0.462–0.123

1 0.754

— 0.469–1.218

0 ⫺1.609 1.659 1.084 —

— ⫺4.418–2.638 ⫺3.085–8.462 0.678–1.729 —

1 0.018 0.000 — 1.214

— 0.008–0.043 0.000–0.002 — 1.135–1.298

0 ⫺0.004

— ⫺0.659–0.668

1 26.843

— 11.977–61.499

KEY: LE ⫽ life expectancy; CI ⫽ confidence interval; ICED ⫽ Index of Coexistent Disease10 (0 ⫽ healthy; 2 ⫽ angina, previous myocardial infarction; 3 ⫽ angina, chronic obstructive pulmonary disease, previous myocardial infarction). * Difference between estimated LE and true (model’s) LE, relative to the reference category. The final parameter estimates are the means from the 1000 regression analyses; the CIs incorporate error from within each regression (usual standard error) and between regressions (because of uncertainty in true LE). For example, the clinicians overestimated LE of the 65-year-old men by 1.285 years, relative to the error in their LE estimates for the 55-year-old patients, but it was not statistically significant.

FIGURE 2. Treatment recommendations (percentage radical prostatectomy) as a function of estimated life expectancy. Size of circle is proportional to number of observations. Some values for life expectancy, such as 5 or 10, were given more frequently than others, such as 7 or 12.

COMMENT One approach to interpreting the results of this study might be the concept of “clinical significance.” Several studies have suggested that 1 year is a moderate-to-large gain in life expectancy.18 –20 Clearly, the average error in the clinicians’ estimates was larger than this. Only one third of estimates were this accurate. Perhaps a more clinically relevant guide to interpretation would be the ability of clinicians to correctly classify patients UROLOGY 60 (2), 2002

FIGURE 3. Treatment recommendations (percentage radiation) as a function of estimated life expectancy. Size of circle is proportional to number of observations. Some values for life expectancy, such as 5 or 10, were given more frequently than others, such as 7 or 12.

around a useful benchmark, such as the “10-year rule.” By this standard, our finding that 82% of clinicians’ estimates correctly classified patients as having life expectancies greater than or less than 10 years suggests a relatively high level of operational accuracy. The results of our study also demonstrated that a 10-year life expectancy was a milestone in the decision to perform radical prostatectomy, although it very interestingly appeared to have been regarded as a minimal threshold, rather than a clear 261

indication for surgery. Life expectancy thresholds were also observed for radiotherapy (5 to 15 years) and watchful waiting (less than 5 years). Although the 10-year rule is frequently cited,1– 4 its origins lie shrouded in mystery. One local clinician described its evidentiary status as “urological dogma,” although our data suggest that radiation oncologists can also fairly lay claim to this canine heuristic. What seems to underlie its acceptance is the belief that because prostate cancer progresses slowly, a prolonged period of expected survival is required before a survival benefit can be observed. The concept of life expectancy is, however, surprisingly subtle.18 Any separation in the survival curves of different treatment groups leads to a reduction in the risk of death and an increase in life expectancy. Thus, use of a fixed time threshold before a treatment is recommended (eg, the 10year rule) suggests that survival curves only diverge after this point. Moreover, in the context of prostate cancer, the 10-year rule incorporates the implicit assumption that gains in life expectancy do not depend on tumor factors (grade, stage, prostate-specific antigen level), when, in fact, patients with high-grade tumors are likely to gain considerably more life years from aggressive treatment than those with lower grade disease.21 Ultimately, the potential gains in life expectancy afforded by treatment, rather than some fixed life expectancy threshold, may be used to guide practice. Gains in life expectancy will reflect age and comorbidity, but also disease factors and efficacy of therapy. Clinician specialty was a strong predictor of treatment recommendation in this study. Similar results have been reported elsewhere.3,22 This is concerning, and suggests that factors other than evidence may be affecting clinical decisions. The response rate of this survey (69%) was quite high, but included only Canadian clinicians. Despite some variation in practice patterns,23 similar guidelines for age and comorbidity are used in Britain and the United States, suggesting that our results are generalizable.1,23–26 One limitation of our study was its format. In a clinical setting, the history, examination, and observation of functional status provide additional elements that cannot readily be captured in scenarios. CONCLUSIONS Our results indicate that clinicians, using unaided clinical judgment, can use age and comorbidity information to predict patients’ life expectancy with a modest degree of overall accuracy, but a high degree of operational accuracy with respect to the effective use of the 10-year rule. The validity 262

of the rule remains to be determined by future decision-modeling studies and randomized trials. REFERENCES 1. Royal College of Radiologists’ Clinical Oncology Information Network, British Association of Urological Surgeons: Guidelines on the management of prostate cancer. BJU Int 84: 987–1014, 1999. 2. Goldenberg SL, Ramsey EW, and Jewett MAS: Prostate cancer: 6. Surgical treatment of localized disease. Can Med Assoc J 159: 1265–1271, 1998. 3. Fowler FJ, Collins MM, Albertsen PC, et al: Comparison of recommendations by urologists and radiation oncologists for treatment of clinically localized prostate cancer. JAMA 283: 3217–3222, 2000. 4. Koch MO, Miller DA, Butler R, et al: Are we selecting the right patients for treatment of localized prostate cancer? Results of an actuarial analysis. Urology 51: 197–202, 1998. 5. Parkes CM: Accuracy of predictions of survival in later stages of cancer. BMJ 2: 29 –31, 1972. 6. Vigano A, Dorgan M, Bruera E, et al: The relative accuracy of the clinical estimation of the duration of life for patients with end of life cancer. Cancer 86: 170 –176, 1999. 7. Christakis NA, and Lamont EB. Extent and determinants of error in doctors’ prognoses in terminally ill patients: prospective cohort study. BMJ 320: 469 – 473, 2000. 8. Llobera J, Esteva M, Rifa J, et al: Terminal cancer: duration and prediction of survival time. Eur J Cancer 36: 2036 – 2043, 2000. 9. Canadian Medical Association: Canadian Medical Directory 1997. Don Mills, Ontario, Canada, Southam Information Products Group, 1997. 10. Greenfield S, Apolone G, McNeil BJ, et al: The importance of coexistent disease in the occurrence of postoperative complications and one-year recovery in patients undergoing total hip replacement. Med Care 31: 141–154, 1993. 11. TreeAge Software. Data 3.0. Williamstown, Massachusetts, 1998. 12. Alibhai SMH, Nam R, Naglie G, et al: The impact of age and co-morbidity in the treatment of clinically localized prostate cancer—a decision analysis. Gerontologist 39(Special issue 1): 287, 1999. 13. Statistics Canada. Life tables, Canada and provinces 1985–1987. Health Rep 2(4 suppl 13): 1–58, 1990. 14. Albertsen PC, Fryback DG, Storer BE, et al: Long-term survival among men with conservatively treated localized prostate cancer. JAMA 274: 626 – 681, 1995. 15. Statistical Package for the Social Sciences for Windows, version 9.0. SPSS Inc., Chicago, 2000. 16. S-Plus 2000 User’s Guide. Data Analysis Products Division, MathSoft, Seattle, 2000. 17. Spiegelhalter DJ, Thomas A, and Best NG: WinBUGS Version 1.2 User Manual. Cambridge, England, MRC Biostatistics Unit, 1999. 18. Naimark D, Naglie G, and Detsky AS: The meaning of life expectancy: what is a clinically significant gain? J Gen Int Med 9: 702–707, 1994. 19. Krahn MD, Naglie G, Naimark D, et al: Primer on medical decision analysis: part 4 —analyzing the model and interpreting the results. Med Decis Making 17: 142–151, 1997. 20. Wright JC, and Weinstein MC: Gains in life expectancy from medical interventions—standardizing data on outcomes. N Engl J Med 339: 380 –386, 1998. 21. Albertsen PC, Hanley JA, Gleason DG, et al: Competing risk analysis of men aged 55 to 74 years at diagnosis managed conservatively for clinically localized prostate cancer. JAMA 280: 975–980, 1998. UROLOGY 60 (2), 2002

22. Moore MJ, O’Sullivan BO, and Tannock IF: How expert physicians would wish to be treated if they had genitourinary cancer. J Clin Oncol 6: 1736 –1745, 1988. 23. Fleshner N, Rakovitch E, and Klotz L: Differences between urologists in the United States and Canada in the approach to prostate cancer. J Urol 163: 1461–1466, 2000. 24. Savage P, Bates C, Abel P, et al: British urological surgery practice: 1. Prostate cancer. Br J Urol 79: 749 –755, 1997.

UROLOGY 60 (2), 2002

25. Donovan JL, Frankel SJ, Faulkner A, et al: Dilemmas in treating early prostate cancer: the evidence and a questionnaire survey of consultant urologists in the United Kingdom. BMJ 318: 299 –300, 1999. 26. Barry MJ, Fowler FJ, Bin L, et al: A nationwide survey of practicing urologists: current management of benign prostatic hyperplasia and clinically localized prostate cancer. J Urol 158: 488 – 492, 1997.

263