170 The Relative Accuracy of the Clinical Estimation of the Duration of Life for Patients with End of Life Cancer Antonio Viganò, M.D.1 Marlene Dorgan, M.L.S.2 Eduardo Bruera, M.D.1 Maria E. Suarez-Almazor, BACKGROUND. Although the prediction of the duration of life of patients with end 3 Ph.D. 1 Division of Palliative Care Medicine, Grey Nuns Hospital, Edmonton, Alberta, Canada. 2 J. W. Scott Health Sciences Library, University of Alberta, Edmonton, Alberta, Canada. 3 Department of Public Health Sciences, University of Alberta, Edmonton, Alberta, Canada. of life cancer most often relies on the clinical estimation of survival (CES) made by the treating physician, the accuracy and practical value of CES remains controversial. METHODS. The authors prospectively evaluated the accuracy of CES in an inception and population-based cohort of 233 cancer patients who were seen at the onset of their terminal phase. They also systematically reviewed the literature on CES in advanced or end-stage cancer patients in MEDLINE, CANCERLIT, and EMBASE data bases, using two search strategies developed by a research librarian. RESULTS. CES had low sensitivity in detecting patients who died within shorter time frames (#2 months), and a tendency to overestimate survival was noted. A moderate correlation was observed between actual survival (AS) and CES (Pearson correlation coefficient 5 0.47, intraclass correlation coefficient 5 0.46, weighted kappa coefficient 5 0.42). CONCLUSIONS. Treating physicians appear to overestimate the duration of life of end of life ill cancer patients, particularly those patients who die early in the terminal phase and who may potentially benefit from earlier participation in palliative care programs. CES should be considered one of many criteria, rather than a unique criterion, by which to choose therapeutic intervention or health care programs for patients in the end of life cancer phase. Cancer 1999;86:170 – 6. © 1999 American Cancer Society. KEYWORDS: prognosis, survival, end of life cancer, life expectancy. D Presented in part at the 12th International Congress on Care of the Terminally Ill, Montreal, Quebec, Canada, September 13–17, 1998. The authors are grateful to Mrs. Brenda Topliss for her invaluable technical assistance. Address for reprints: Maria E. Suarez-Almazor, Ph.D., 214 Heritage Medical Research Centre, Faculty of Medicine, University of Alberta, Edmonton, Alberta T6G 2S2, Canada. Received September 14, 1998; revision received January 25, 1999; accepted February 16, 1999. © 1999 American Cancer Society espite major medical advances, the rate of incidence and the number of cancer deaths have increased substantially in recent years. The number of people diagnosed with cancer in Canada has increased by 30% in the last decade, and it is estimated that in developed countries approximately 30% of the population alive today will develop cancer in their lifetimes.1 For approximately 50% of patients diagnosed with cancer, active treatments (aimed primarily at a cure or prolongation of life) become ineffective at a certain point.2 Although the definition of life expectancy for these patients often relies on the clinical estimation of survival (CES), it is still unclear whether this information should be used in the counseling of patients and families and/or in establishing patients’ eligibility for tumordirected therapies or palliative care programs. In the U.S., a 1993 report from the National Hospice Organization showed that over 50% of patients with end of life cancer were not given access to hospice services3 or were referred too late in the course of their illness to take full advantage of the support provided by hospice programs.4 Overly Survival Prediction and End of Life Cancer/Viganò et al. optimistic CES made by different health care providers has been suggested to be among the major determinants of these late referrals.5 On the other hand, some studies have shown the accuracy of clinical predictions of survival to be similar6 or superior7 to predictions based on the assessment of characteristics of major prognostic importance, such as performance status. Heterogeneity among studies could perhaps explain some of these differences. Some authors have indicated that the magnitude of the interaction between CES and actual survival may vary according to the length of follow-up or the median survival of the sample, as has been shown for other survival predictors.8 Other studies have suggested that the difficulty in predicting terminal survival may be that death among cancer patients is due to many essentially unpredictable events.9 –11 To define better the role of physicians’ prediction in the management of end of life cancer patients, we prospectively evaluated the accuracy of their CES in a population-based cohort of patients who were seen at the onset of their terminal phase. We also systematically reviewed the available literature on this subject. METHODS An inception cohort of sequentially accrued patients with advanced cancer was assembled at the Cross Cancer Institute (CCI) in Edmonton, Alberta, Canada. The Institute is the only referral center for oncologic treatment in northern Alberta, and its catchment area has a population of approximately 1.5 million. During the accrual period, 270 patients who were seen as outpatients or admitted as inpatients were potentially eligible for the study. Patients were eligible if they were older than 18 years and had terminal cancer of the lung, breast, gastrointestinal system, or prostate. Their diseases were defined as terminal when further attempts to arrest or control progression were deemed unavailable.12 Patients were considered eligible for the study if they entered the terminal phase within 30 days of their possible accrual. Of the 270 patients who met the inclusion criteria, 248 (92%) agreed to participate in the study. Patients were given an initial assessment followed by monthly telephone interviews throughout the course of their disease until death or the end of the study. At the time of the initial assessment, the treating oncologist was asked to provide a CES in weeks or months. The CES was recorded for 233 patients. Statistical Analysis The accuracy of CES was compared with the actual survival of patients using the following methods: 171 1. The difference between CES and actual survival was calculated for each patient. One-way analysis of variance (ANOVA) was performed to compare the mean values of these differences among different oncologists. The Bonferroni test was applied for multiple comparison procedures.13 To determine whether prediction errors were associated with particular physicians or with patient or disease characteristics of clinical and prognostic value, logistic regression analyses were performed. The dependent variable was dichotomous: accurate (within 1 month of error) versus inaccurate (error .1 month). Independent variables were age, gender, type of disease, type of treatments received during and after study enrollment, performance status, and source of referral. 2. The relation between actual survival and CES was assessed with the Pearson correlation coefficient (r) and the intraclass correlation coefficient (ICC). The ICC is a reliability coefficient calculated from an ANOVA. The ICC reflects the degrees of both correspondence and agreement between ratings, adjusting for the role of chance in the agreement. 3. Both CES and actual survival were categorized into 1) #2 months, 2) 2– 6 months, and 3) .6 months. These categorizations are commonly used both in Canada and in the U.S. to provide access to government-funded hospices and regional and hospital-based palliative care units. The median length of stay in hospices in the U.S. is 29 days; consequently, half of patients are referred to hospices with less than 1 month to live.14 The sensitivity, specificity, positive (PPV) and negative predictive values (NPV), and accuracy of CES were estimated as follows: Sensitivity was calculated as the percentage of patients in a specific survival category who were correctly predicted— that is, the number correctly predicted within that category divided by the total number of patients whose actual survival was within that category. Specificity was calculated as the number of patients correctly classified as not belonging to a specific survival category divided by the number of patients wrongly classified as belonging to a specific survival category plus the number of patients correctly classified as not belonging to a specific survival category. The PPV was calculated as the percentage of patients, of those predicted by the physician to be within a specific survival category, who were correctly predicted—that is, the number correctly predicted within that category divided by the total number of patients whose survival was predicted to be in that category. The NPV was calculated as the percentage of 172 CANCER July 1, 1999 / Volume 86 / Number 1 patients who were correctly predicted not to belong to a specific survival category—that is, the number correctly predicted not to be in that survival category divided by the total number of patients whose survival was predicted to be in that category. Accuracy represents the percentage of patients correctly classified by physicians within the categories. These cutoffs were also used to calculate an ordinal measure of agreement (weighted kappa, k).15 Kappa is the ratio of the number of observed nonchance agreements to the number of possible nonchance agreements. Incremental weights were assigned to disagreement cells in the agreement matrix, according to the formula w 5(r1–r2)2, where w is the assigned weight and r1–r2 represents the deviation from agreement between CES and actual survival for each cell in the agreement matrix. For patients who were predicted to live between 2 and 6 months but actually lived either #2 months or .6 months, the observed and expected frequencies were multiplied by a w of 1, whereas for patients who were predicted to live either #2 months or .6 months but actually lived either .6 months or #2 months, respectively, w was calculated to be 4.15 Statistical significance was set at P # 0.05 (twotailed). All analysis were carried out using the SPSS 6.1 statistical software package.16 Publications included in the review of the literature on CES for advanced or end-stage cancer patients were initially identified through MEDLINE, CANCERLIT, and EMBASE, using two search strategies (which will be provided by the authors on request) developed by a research librarian. The years 1966 –1998 for MEDLINE, the years 1983–1998 for CANCERLIT, and the years 1980 –1998 for EMBASE were respectively searched. We also examined the reference lists from selected articles. Data reviewed and extracted from selected articles included sample size, type of primary tumor, length of survival, type of study, whether the cohort was selected at the inception of the terminal phase, correlation coefficients (r) between actual survival and CES, and other relevant outcomes of the studies. The above design features are thought to be generally relevant to the quality of studies on prognosis; the correlation coefficient was one of few elements to allow quantitative comparisons among the studies.17–19 RESULTS Table 1 shows selected demographic and clinical characteristics of the 248 patients originally included in the study. At the time of analysis (April 1998), 225 patients TABLE 1 Selected Demographic and Clinical Characteristics of 248 Patients Characteristic No. Median age, yrs (range) Status Dead Alive Lost to follow-up Gender Male Female Race White Nonwhite Primary site Breast Gastrointestinal Lung Prostate Tumor-directed treatmentsa No treatment Chemotherapy Radiotherapy Hormonotherapy 62 (29–92) % 225 23 0 91 9 0 103 145 42 58 229 19 92 8 70 80 77 21 28 32 31 9 84 34 113 17 34 14 45 7 a Tumor-directed treatment was administered concurrently or after study accrual of patients, with palliative intent. (91.0%) had died, 23 (9.0%) were alive, and no patient was lost to follow-up. The estimated median survival time for the group overall was 15.3 weeks. The difference between CES and actual survival was calculated for the 210 patients who were known to be dead and who had had their survival predicted. Fifteen patients had had their date of death recorded but had not had their survival predicted. The accuracy of CES is presented in Figure 1, which shows the differences between CES and actual survival. The median difference between CES and actual survival was 1.1 month (range, 612 months) in the optimistic direction. Fifty-two patients (25%) had had their survival correctly predicted by their treating oncologist to within 1 month, 49 (23%) had had their survival underestimated, and 109 (52%) had had their life expectancy overestimated. The Pearson correlation coefficient between actual survival and CES was 0.47 (P , 0.0001) and the intraclass correlation coefficients were 0.46 and 0.47, respectively, for random and fixed models. No differences were noted in the accuracy of prediction by different oncologists; therefore, there was no need to stratify the analysis further according to physician. This was further supported by the result of a logistic regression analysis that failed to demonstrate any significant associations between inaccurate prediction and source of referral (medical oncologists vs. Survival Prediction and End of Life Cancer/Viganò et al. 173 FIGURE 1. Difference between the clinical estimation of survival and the actual survival. TABLE 2 Overall Comparison of Clinical Estimation of Survival (CES) and Actual Survival Actual Survival Clinical estimation of survivala <2 mos 2–6 mos >6 mos Total #2 mos 2–6 mos .6 mos Total 23 (31%) 42 (57%) 9 (12%) 74 5 (6%) 53 (68%) 20 (26%) 78 3 (4%) 34 (42%) 44 (54%) 81 31 129 73 233 Clinical estimation of survivala Sensitivity Specificity Positive predictive value (PPV) Negative predictive value (NPV) #2 mos 2–6 mos .6 mos 31% 68% 54% 95% 51% 81% 74% 41% 60% 75% 76% 77% a Accuracy 5 52%; weighted kappa 5 0.42. radiotherapy oncologists), age and gender of patients (male vs. female, #65 vs. .65 years), type of disease (lung vs. nonlung), treatment choice for antitumor therapy in the terminal phase (no treatment or treatment discontinued vs. treatment initiated or continued), and performance status (bedridden vs. ambulatory). The categorization of both actual survival and CES allowed the inclusion of patients who were still alive at the date of study closure. Two hundred thirty-three patients were included in one of the above categories (Table 2). Almost half of the patients were not assigned to the correct survival categories, and the weighted k only showed moderate agreement between CES and actual survival (k 5 0.42). In particular, low sensitivity of CES was observed for patients who died within 2 months, suggesting that physicians had some diffi- culty identifying patients with short survival. Sensitivity was also low (54%) for patients who survived longer than 6 months. On the contrary, specificity was higher for shorter and longer survival. These findings show that oncologists were more likely to predict survival periods of between 2 and 6 months, often inaccurately. Positive predictive values in this context reflect the probability that a given CES is accurate; they were 74% for #2 months, 41% for 2– 6 months, and 60% for .6 months. When physicians predicted a short survival period, they were generally correct (PPV 74%), but they only predicted this for a small proportion of those patients who actually survived ,2 months (sensitivity 31%, overestimation 79%). Several hundred publications were identified through MEDLINE, CANCERLIT, and EMBASE searches. Of these, 11 studies were related to CES and 174 CANCER July 1, 1999 / Volume 86 / Number 1 TABLE 3 Methodologic Features and Relevant Outcomes of the 11 Studies Included in the Review Histology Median survival (weeks) Inception cohort Sampling frame Pearson r between As and CES 168 178 All Lung NRa 6–22 No No Hospice Clinical trial (range) 0.24–0.28 NR Evans, 19856 42 All NR No Home care 0.44 (initial CES) 0.58 (final CES) Heyse-Moore, 198722 Forster, 198823 50 108 All All 2 3.5 No No Hospice Hospice .23 (range) 0.30–0.40 Addington-Hall, 199024 Bruera, 199225 469 All 17.5 No Hospital NR 47 All 4b No Palliative care unit NR Maltoni, 19947 100 All 5 No Home care 0.51 Maltoni, 199526 540 All 4.6 No Home care NR Muers, 199627 196 Lung 18 No Hospital NR Oxenham, 199828 30 All 2.4 No Hospice (range) 0.31–0.72 (initial CES) (range) 0.61– 0.98 (final CES) Study (by first author) No. of patients Parkes, 197220 Scotto, 197221 Other relevant outcomes 87% of errors were optimistic. Physician could clearly identify patients with prognoses ,2 mos or .6 mos. Equal percentage of over estimation and underestimation. Predictions tended to be overoptimistic. Karnofsky performance index gave a closer correlation with AS than CES. Final CES better than initial ones. Most predictions were over-estimated. High variation in performances of different health professionals. Poor correlation coefficients among the estimates made by different health professionals. Serious errors tended to overestimate survival. CES were fairly accurate within a 3-mo boundary. Doctors and nurses were as likely to over estimate as to underestimate life expectancies .1 or ,1 yr. PPV 5 0.86, NPV 5 0.64. Accuracy 5 0.68 for CES of patients who died/survived within/beyond 4 weeks. Difference in AS-CES 5 1 wk. For 62% of patients survival was overestimated and for 37% underestimated. CES performed better than KPS assessment (r 5 0.37). Better correlation AS-CES according to level of physician experience. CES remained statistically significant in multivariate analysis. For 36% of patients survival time was underestimated, for 55% overestimated. Sensitivity 5 52%, 73%, 38%; PPV 5 80%, 50%, 49%; accuracy 5 57%, for predictions of #3, 3–9 and .9 mos. CES could differentiate further prognostic groups in a multivariate model. Predictions made within 1 week of death more accurate than initial CES. Auxiliary nursing provided better prognosticators than senior team members when patients were near death. AS: actual survival; CES: clinical estimation of survival; N: sample size; MS: median survival; PCU: palliative care unit; KPS: Karnofsky performance status assessment; NR: not reported; r: correlation coefficient; PPV: positive predictive value; NPV: negative predictive value. a Eighty-three percent of the sample died within 12 weeks. b Mean survival. were reviewed6,7,20 –28 (Table 3). These studies reported median survival periods ranging from 2.4 to 22 weeks. Only two studies examined nonsmall cell lung carcinoma as specific primary tumor,21,27 whereas the remaining considered any end stage malignancies. Cohorts were not identified at the inception of the terminal phase in any of the studies, and only 224,27 of 11 included population-based samples. In 6 of 11 studies in which this information was available, the correlation coefficients between actual survival and CES were in the poor-to-moderate range. Seven studies clearly reported a tendency to overestimate survival, whereas in another two,21,24 health professionals appeared as likely to overestimate as to underestimate life expectancies. Similarly to our findings, in which health professionals actually predicted a short survival period, they were fairly accurate, with moderate-tohigh PPV.25,27 Sensitivity, however, was low, because they only predicted a short survival period in a proportion of those who actually died within the first 2–3 Survival Prediction and End of Life Cancer/Viganò et al. months. Poor correlation among the estimates of different health professionals was reported in one study.23 In another, CES appeared to be independent of other survival indicators in multivariate analyses.26,27 These findings suggest that CES does not rely on “common observations” and that it is highly dependent on the individual perceptions of the person making the determination. The level of personal knowledge of the patient appears to improve the prediction.6,28 It is still unclear whether CES is equally improved by the seniority level of health professionals, although seniority did not appear to have an impact in our study. DISCUSSION The objective of this study was to evaluate the accuracy of CES in patients with end of life cancer. Although some other studies have also assessed clinical predictions of survival, they have generally been based on patients admitted to hospices or receiving palliative care in selected settings. We included in our study a population-based cohort of patients consecutively accrued at a specific inception point, defined as the beginning of the terminal phase as determined by the treating oncologist. As such, our sample is more representative of patients with end of life cancer in the population at large than are some previous studies. A moderate correlation was observed between actual survival and CES, with values of approximately 0.50 for the bivariate correlation coefficient, the ICC, and the weighted k. These last two coefficients take into account agreement by chance alone and are more appropriate than the simple correlation coefficients used in previous studies.18 Only 25% of the physicians’ predictions of survival were accurate within 1 month. Two-thirds of the incorrect CES overestimated survival, particularly for those patients who had shorter survival. Overestimation of survival has also been reported by others.25–27 Survival was categorized into 3 groups: #2 months, 2– 6 months, and .6 months. These categories were based on cutoff values applied for the provision of palliative care services in some settings. Using these broad categories, the accuracy of the CES improved to 52%. These values are also consistent with those reported in other studies that classified survival into similar groups.25,27 In general, when physicians predicted a short survival period, they were correct (PPV 74% for CES #2 months). However, they only predicted this short survival period for one-third of those who actually died within 2 months of entering the terminal phase. A concern regarding overestimation of survival is that some patients may be denied prompt access to palliative care 175 if there is an erroneous expectation of longer survival, because some health care programs are only accessible to patients who are expected to live for less than 2 months. In our study, physicians were also more likely to be accurate when predicting survival periods longer than 6 months. Nevertheless, for 40% of the patients who actually lived longer than 6 months, the treating physician had predicted a shorter survival period. Most often, physicians expected the survival of terminally ill patients to be within 2 and 6 months, when, in fact, one-third survived 2 months or less, another third 2–6, months and the remaining one-third longer than 6 months. These findings suggest that CES on its own is not an accurate parameter for a quantitative prognosis. However, a few studies that examined the potential value of CES when included in prognostic models based on patient characteristics have found that CES remains an independent survival predictor, adding some information not otherwise collected or measured.26,27 It has been suggested that CES relies on a “subconscious appraisal.”29 Our findings and review of the literature confirm that the performance of CES is independent of the level of seniority among health professionals24,28 and of other patient and disease characteristics that have been found to be of significance to the prognoses of terminally ill cancer patients.26,27 The major problem with CES appears to be its lack of sensitivity in identifying patients with poorer survival for whom an adequate delivery of palliative care would be the most appropriate. An extensive review of patients’ medical charts, as performed by Forster et al.,23 or the “subconscious appraisal” that seems to have characterized CES in our study and other studies,26,27,28 do not work well by themselves. It is possible that a better knowledge of patients, in combination with some other possible predictors of survival, may improve prediction, as has been suggested in some of the reviewed publications.6,25,28 This study and our review of the literature show that health professionals are often inaccurate in their survival predictions for patients with end of life cancer, and that they tend to overestimate life expectancy, particularly for patients who die soon after entering the terminal phase. These patients could perhaps benefit from being involved in palliative care programs at an earlier stage. Better estimation of survival may be achieved by adding some clinical indicators of prognostic value (i.e., performance status, nutritional parameters, etc.) to a prognostic model. On its own, CES appears to be inadequate for making decisions in relation to the care of the terminally ill. It should be considered one of many criteria, rather than the sole benchmark, by which to choose a therapeutic inter- 176 CANCER July 1, 1999 / Volume 86 / Number 1 vention or health care program for cancer patients in the terminal phase. REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. Fraumeni JF, Devesa SS, Hoover R, Kinlen L. Epidemiology of cancer. In: De Vita V, Hellman S, Rosemberg S, editors. Cancer principles and practice of oncology. 4th edition. Philadelphia: J. B. Lippincott, 1994:150 – 82. WHO Expert Committee. Cancer pain relief and palliative care. Geneva: World Health Organization, 1990. National Hospice Organization (NHO) Newsline. 1992 stats show continued growth in programs and patients. NHO Newsline 1993;3:1–2. Demer C, Johnston-Anderson AV, Tobin R, Cohen MH, et al. Cost of hospice care: late versus early entry [abstract]. Proc Am Soc Clin Oncol 1992;11:392. Kinzbrunner BM. Ethical dilemmas in hospice and palliative care. Support Care Cancer 1995;3:28 –36. Evans C, McCarthy M. Prognostic uncertainty in terminal care: can the Karnofsky index help? Lancet 1985,25,1204 – 6. Maltoni M, Nanni O, Derni S, Innocenti MP, Fabbri L, Riva N, et al. Clinical prediction of survival is more accurate than the Karnofsky performance status in estimating life span of the terminally ill cancer patients. Eur J Cancer 1994;30A:6; 764 – 6. Byar DP, Mantel N, Hankey B. Predicting survival in terminal cancer. Br Med J 1973;1:611. Den Daas N. Estimating length of survival in end-stage cancer: a review of the literature. J Pain Symptom Manage 1995;10:548 –55. Klastersky J, Daneau D, Verhest A. Causes of death in patients with cancer. Eur J Cancer 1972;8:149 –54. Inagaki J, Rodriguez V, Bodey G. Causes of death in cancer patients. Cancer 1973;2:568 –73. Saunders CM. Appropriate treatment, appropriate death. In: Saunders C, editor. The management of terminal disease. London: Arnold, 1978. Portney LG, Watkins MP. Analysis of variance. In: Portney LG, Watkins MP. Foundations of clinical research: applications to practice. Norwalk: Appleton & Lange, 1993:373–95. Christakis NA. Timing of referral of terminally ill patients to an outpatient hospice. J Gen Intern Med 1994;9:314 –9. Portney LG, Watkins MP. Statistical measures of reliability. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. In: Portney LG, Watkins MP. Foundations of clinical research: applications to practice. Norwalk: Appleton & Lange, 1993:505–27. SPSS Inc. SPSS for Windows. Base system user’s guide. Release 6.0. Chicago: SPSS Inc.,1993:350. Sackett DL, Haynes RB, Guyatt GH, Tugwell P. Making a prognosis. In: Sackett DL, Haynes RB, Guyatt GH, Tugwell P, editor. Clinical epidemiology: a basic science for clinical medicine. 2nd edition. Boston: Little, Brown, 1991. Laupacis A, Wells G, Richardson WS, Tugwell P, for the Evidence-Based Medicine Working Group. Users’ guides to the medical literature. V. How to use an article about prognosis. JAMA 1994;272:234 –7. Sackett DL, Richardson WS, Rosemberg W, Haynes RB. Evidence-based medicine. New York: Churchill Livingstone, 1997:85–90. Parkes CM. Accuracy of predictions of survival in later stages of cancer. Br Med J 1972;2:29 –31. Scotto J, Schneiderman MA. Predicting survival in terminal cancer. Br Med J 1972;4:50. Heyse-Moore LH, Johnson-Bell VE. Can doctors accurately predict the life expectancy of patients with terminal cancer? Palliat Med 1987;1:165– 6. Forster LE, Lynn J. Predicting life span for applicants to inpatient hospice. Arch Intern Med 1988:148;2540 –3. Addington-Hall JM, MacDonald LD, Anderson HR. Can the Spitzer Quality of Life Index help to reduce prognostic uncertainty in terminal care? Br J Cancer 1990,62:695–9. Bruera E, Miller MJ, Kuehn N, MacEachern T, Hanson J. Estimated survival of patients admitted to a palliative care unit: a prospective study. J Pain Symptom Manage 1992;7: 82– 6. Maltoni M, Pirovano M, Scarpi E, Marinari M, Indelli M, Arnoldi E, et al. Prediction of survival of patients terminally ill with cancer. Cancer 1995;75:2613–22. Muers MF, Shelvin P, Brown J. Prognosis in lung cancer: physicians’ opinion compared with outcome and a predictive model. Thorax 1996;51:894 –902. Oxenham D, Cornbleet MA. Accuracy of prediction of survival by different professional groups in a hospice. Palliat Med 1998;12:117– 8. Holland JF. Prediction of time of death in patients with advanced cancer. AnnN Y Acad Sci 1969;164:678 – 86.