INTRODUCTION

Symptomatic metastatic hormone-refractory prostate cancer (HRPC) patients are usually considered as a homogeneous group whose prognosis remains poor, with a median overall survival of 10 months.¹ However, there is evidence that within this category of patients, subgroups have differing prognostic profiles and experience different durations of survival.^2-13

To gain insight into the prognosis of those patients is essential to the physician's daily need to inform and support the patient as effectively as possible and is also of great importance to the patient himself.¹⁴ Patients with a short life expectancy are generally less willing to undergo extensive treatment with possible severe side effects.¹⁴ Finally, prognostic factors and expected survival are useful for designing clinical trials because they influence the choice of therapy and the patient selection and their knowledge can increase the statistical power of comparisons by reduction of the model variability.¹⁵ To be of use in daily practice, prognostic models should be composed of factors that are easily assessable in clinical practice and should provide precise predictions.

A number of published reports address prognostic factors for survival of (bone) metastatic HRPC patients^2-13 (Table 1). The prognostic factors identified so far are biochemical (baseline prostate-specific antigen [PSA], lactate dehydrogenase [LDH], alkaline phosphatase, acid phosphatase, hemoglobin, and creatinine),^2-13 objective (number of bone metastases, duration of response to primary hormonal treatment, Gleason sum, age, and weight),^{2,5-7,9,11,13} or subjective (performance status^2,5-8,11-13 and disease-related symptoms^2,4,7,8). Among the disease-related symptoms, pain,^2,8 anorexia and appetite loss,^2,8 obstructive voiding problems,² fatigue,⁴ and impairment of physical functioning^7,8 have been identified as prognostic factors in four studies.

There are theoretical arguments to support the hypothesis that health-related quality-of-life (HRQOL) parameters may be important prognostic factors. For example, there are arguments that patients' overall HRQOL is influenced by psychosocial factors, such as stress, social support, emotional expression, and coping strategies,^16-18 and that these psychosocial factors seemed to be of prognostic value for survival in a few studies in cancer.^19-21

HRQOL factors were reported to be independent prognostic factors in four HRPC studies.^2,4,7,8 However, HRQOL was measured by means of validated patient self-assessment questionnaires in only one study.⁸ In the other three studies, the symptoms were assessed either by the doctor² or by means of ad hoc questionnaires.^4,7 Because the use of validated^22,23 patient self-assessment^24,25 questionnaires is a prerequisite for measuring HRQOL reliably, their relevance as prognostic factors remains to be confirmed.

Since 1990, the European Organisation for Research and Treatment of Cancer (EORTC) Genitourinary Group has conducted three randomized clinical trials in patients with (bone) metastatic HRPC. HRQOL was measured in all studies by means of the EORTC Quality of Life Questionnaire C30 (QLQ-C30; version 1.0) before the start of the treatment and during follow-up. The HRQOL data from these studies have been used together with previously identified clinical and biochemical parameters to identify independent prognostic factors for overall survival. The predictive ability of the resulting prognostic model is compared to that of models based on clinical and biochemical factors only.

PATIENTS AND METHODS

The data from three randomized EORTC HRPC trials were used (EORTC 30903,²⁶ EORTC 30921,²⁷ and EORTC 30944²⁸). The trials were conducted in accordance with the Helsinki Declaration and were approved by the local institutions' ethical committees. All patients provided informed consent before randomization. Two hundred one patients with painful metastatic HRPC were entered onto the phase III trial 30903 and randomly assigned to treatment with prednisone (5 mg qid orally [PO]) or flutamide (250 mg tid PO). Two hundred five patients with similar characteristics were randomly assigned in the phase III trial 30921 to receive either one intravenous injection of 150 MBq (4 mCi) of strontium-89 chloride or palliative local radiotherapy with the aim to control pain from osseous metastases. Last, 92 patients were entered onto the phase II trial 30944 and randomly assigned to receive chemotherapy consisting of estramustine phosphate (10 mg/kg/d PO) with or without vinblastine (4 mg/m² intravenously, weekly for 6 weeks every 8 weeks).

Baseline HRQOL assessment was obtained within 1 week of randomization before the start of the treatment. It was available for 391 (78.5%) of the 498 patients. The group without baseline HRQOL had to be excluded from the analysis. Because the groups with and without baseline HRQOL had similar baseline characteristics (Table 2) and overall survival (Fig 1), the analyzed subset can be regarded as being representative of the whole.

Of the 391 patients included in the analysis, 371 patients (94.9%) have died; 326 of the patients (87.9%) died of prostate cancer. The median survival was 10.4 months (95% CI, 9.2 to 11.5 months), and the 1-year survival rate was 42.7% (95% CI, 37.8% to 47.6%).

The EORTC QLQ-C30 version 1.0 was used to assess the HRQOL in the three trials. This validated measure is one of the most frequently used in cancer clinical trials in Europe and has been granted for use in over 3,000 clinical trials.²⁹ It consists of 30 items addressing the functioning and symptoms of cancer patients, as well as their general well-being. From the 30 questions, six multi-item function scales are built (physical functioning, role functioning, emotional functioning, cognitive functioning, social functioning, and global health status/quality of life), along with eight symptom scales (three multi-item scales addressing fatigue, nausea and vomiting, and pain and five single items addressing dyspnea, insomnia, appetite loss, constipation, and diarrhea) and one single-item scale assessing financial difficulties. The scales were transformed linearly to 0 to 100 range according to the EORTC scoring guidelines,³⁰ so that a higher scale score represents a better level of functioning, whereas for symptom scales, a higher score indicates more symptoms or problems.

The 15 standard scales of the EORTC QLQ-C30 were used in the analysis together with nine clinical or biochemical items that were recorded at baseline in the three trials. These included the age of the patient, whether the patient had an orchiectomy or had previously received luteinizing hormone-releasing hormone analogs, the number of hot spots on the bone scan (0, < 5, 5-15, or > 15/superscan), the initial pain score (0 = no analgesics, 1 = occasional use of non-narcotic analgesics, 2 = regular use of non-narcotic analgesics, 3 = occasional use of narcotic analgesics, and 4 = regular use of narcotic analgesics), the baseline WHO performance status, alkaline phosphatase level (WHO grade), hemoglobin level (WHO grade), and baseline PSA (as multiple of the upper normal limit). Other laboratory measurements (bilirubin, serum creatinine, and blood cell counts) were normal in the vast majority of the patients and were, therefore, of no use to this analysis. The information on visceral and lymph node involvement was not collected in all trials and could not be used for modeling.

The distribution of the discrete variables was assessed, and categories with small numbers (< 25) were pooled together. The quality-of-life scales were dichotomized at the median level. The continuous variables of age and PSA were categorized with cut points at the first and third quartiles. The distribution of the 24 resulting variables is presented in Table 3.

The Cox proportional hazards model stratified for trial and treatment was used.³¹ Preselection for entry of factors into the model-building strategy required significance at the α = .05 level in the univariate analysis. A backward multivariate selection procedure was then applied.³² Internal model validation was performed by the bootstrap resampling technique with 500 bootstrap samples,³³ providing a bias-corrected estimate of the area under the receiver operating characteristic curve (ROC AUC). The latter measures the predictive discrimination of the model (a value of 0.5 indicates no discrimination and a value of 1 indicates perfect discrimination).³³ The model validation was performed using an S-plus (version 3.3; Statistical Sciences, Seattle, WA) function that is available in the public domain.³⁴

RESULTS

The results of the univariate analysis are listed in Table 4. All variables except previous orchiectomy, previous use of luteinizing hormone-releasing hormone, cognitive functioning, diarrhea, dyspnea, pain, and financial difficulties were selected for inclusion in the initial step of the model selection procedure.

At the end of the model selection process, only five variables were retained as independent prognostic factors; these included initial performance status, bone scan result, hemoglobin level, appetite loss, and insomnia (Table 5). Age was last to drop out of the model, with P = .0635. Low hemoglobin level (< 11 g/dL) was associated with a 90% increase in the risk of death. This five-factor model was the most frequently selected model during the bootstrap validation. The bias-corrected ROC AUC for the final model was 0.65. The results were similar when hemoglobin, (logarithm of) alkaline phosphatase, and (logarithm of) PSA were considered as continuous variables in the model. The survival probabilities for groups of patients defined by the tertiles of this prognostic model are displayed in Figure 2.

To assess the increase in predictive ability of our model in comparison to a model that does not include the HRQOL factors, the model-building process was repeated using only the nine clinical and biochemical factors. Only the three factors of WHO performance status, bone scan result, and hemoglobin were retained in the model (Table 6). Age was again borderline significant (P = .0692). The bias-corrected ROC AUC for this model was 0.63. The bias-corrected ROC AUC for a model using only HRQOL factors was 0.59.

DISCUSSION

In this multivariate analysis of overall survival in 391 (bone) metastatic HRPC patients, only the symptom items of appetite loss and insomnia from the EORTC QLQ-C30 were retained as independent prognostic factors of overall survival, in addition to three clinical or biochemical factors (the hemoglobin level, the number of bone metastases on the initial bone scan, and the initial WHO performance status). The factor WHO performance status had a P value of .0322 only, and the group with a performance status of 1 did not significantly differ from the relatively small group (13.3%) with a performance status of 0. None of the HRQOL function scales were selected in the multivariate model. The model evaluation requires that a bone scan, a blood sample, and the completed EORTC QLQ-C30 questionnaire be available. The ROC AUC for this model is 0.65. This value means that, with these five factors, one is able to predict which of two patients will live longer in 65% of the cases.

Of the two HRQOL factors identified as independent prognostic factors in our analysis, only appetite loss was identified earlier as an independent prognostic factor.⁸ Several factors that proved to be of independent prognostic value in earlier prostate cancer studies, such as physical functioning,^7,8 pain,^2,8 and fatigue,⁴ were not identified as such in our analysis. Neither was global health status/quality of life, a factor that was identified as an important prognostic factor in more general cancer patient populations.^35,36 In our data, physical functioning and role functioning were substantially correlated with WHO performance status. Each became statistically significant whenever WHO performance status was forced out of the model. The model validation, however, retained the model with WHO performance status as providing the best prediction. It is surprising that, in our study, global health status/quality of life was not statistically significant even in absence of the factor of WHO performance status. Fatigue was also associated with appetite loss and insomnia; however, this factor was not statistically significant even in absence of the other two factors.

This evaluation confirms that, in HRPC, there is an association between the duration of survival with HRPC and the patient's subjective assessment of their well-being. The statistical significance of HRQOL factors in multivariate prognostic models indicates that these factors are useful in predicting overall survival on average over groups of patients.

However, HRQOL assessments by means of questionnaires are not always easy to obtain. They require extra efforts for the patient and add an extra burden on the shoulders of the clinical or nursing staff at the hospital.^17,19 Therefore, before considering a prognostic model including such factors for routine use in clinical practice, it is worthwhile to assess the amount of improvement to the accuracy of the prediction of the survival of individual patients that such factors bring when they are added to prediction models based on clinical factors only.^15,37

Smaletz et al¹² and Halabi et al¹³ developed risk scores by means of nomograms on two large databases of 409 and 1,101 HRPC patients, respectively. In each, the nomograms were built on the basis of seven clinical and biochemical factors and did not include HRQOL. Smaletz et al¹² used Karnofsky performance status, hemoglobin, alkaline phosphatase, albumin, LDH, baseline PSA, and age. PSA and age were not statistically significant predictors of survival in the multivariate analysis. Internal and external validation of their nomogram achieved a ROC AUC of 0.71 and 0.67, respectively. In Halabi et al,¹³ the nomogram used the presence of visceral disease, Gleason sum, Eastern Cooperative Oncology Group performance status, hemoglobin, alkaline phosphatase, LDH, and baseline PSA. Visceral disease was not statistically significant in the multivariate analysis. The ROC AUC for the model of Halabi et al was 0.68 on the validation set. It is of note that the patient population in Halabi et al had a somewhat better prognosis than our study population, with a median survival of 13 months, and a better performance status. Of the factors used in those two models, ours includes only WHO performance status and hemoglobin level. Alkaline phosphatase and PSA were not statistically significant in our multivariate model. However, when the factor of bone involvement was forced out of the model, alkaline phosphatase was significant and accounted for about half the effect assigned to the bone scan in our model. Albumin and LDH, visceral disease, and Gleason score were not collected in the EORTC studies. Without these variables, we could not assess directly the extra precision that the addition of HRQOL would bring when added to the models of Halabi et al or Smaletz et al. Therefore, we compared the ROC AUC of the five-factor model containing HRQOL factors with that of the best prediction model that could be developed on our database when using only the clinical and biochemical factors. This simplified model included only WHO performance status, hemoglobin, and the number of hot spots on the bone scan and achieved a ROC AUC of 0.63. This was lower than the ROC AUCs achieved by the two nomograms, which could be anticipated from important factors in the models of Halabi et al and Smaletz et al not being available in the EORTC database. Nevertheless, it demonstrates that the addition of the HRQOL component to the three-factor model adds only 1% to the model predictive ability. This is very little and is less than what is obtained by adding clinical information such as Gleason sum, alkaline phosphatase, or LDH in the two nomograms.

From this, we conclude that HRQOL factors, even if they have been shown to be independent prognostic factors in several studies including ours, are not particularly useful for predicting the overall survival of individual HRPC patients once clinical and biochemical factors are taken into account. Given the extra burden they require and the little extra precision they bring to these predictions, we suggest that models based on clinical and biochemical factors should be used whenever all the factors are available from routine practice. When they are not available, only a rough prediction may be obtained on the basis of HRQOL factors only. However, our study shows that other prognostic factors should be investigated to further increase the predictive ability of the existing prediction models proposed by Halabi et al¹³ and Smaletz et al.¹² In that search, the statistical significance of the new factors should not be the only focus, and there must be a careful assessment of the predictive ability of the models and of the extra precision to the predictions that is achieved when the new factors are added to existing models.

Appendix

The following is a list of participants in the trials and their institutions: Trial 30921: Newcastle General Hospital, Newcastle, United Kingdom (Dr J.T. Roberts); Sint Radboud University Hospital, Nijmegen, the Netherlands (Prof G. Oosterhof); Academisch Medisch Centrum, Amsterdam, the Netherlands (Dr Th.M. de Reijke); Rigshospitalet, Copenhagen, Denmark (Dr S.A. Engleholm); Antoni Vanleewenhookhuis, Amsterdam, the Netherlands (Dr S. Horenblas); Aarhus Kommunehospital, Aarhus, Denmark (Dr Nielsen); Ospedale di Circolo e Fundacione Macchi, Varese, Italy (Prof A. Bono); Onze Lieve Vrouw Gasthuis, Amsterdam, the Netherlands (Dr P.P.M. Karthaus); Universitair Ziekenhuis Gent, Gent, Belgium (Dr W. Oosterlinck); Curie Memorial Center, Warsaw, Poland (Prof Madej); Universitair Ziekenhuis Leuven, Leuven, Belgium (Prof H. van Poppel); Allgemeine Ziekenhuis Groningen, Groningen, the Netherlands (Prof H.J.A. Mensink); San Camillo and Forlanini Hospitals, Roma, Italy (Dr C. Sternberg); Bosch Medicentrum, s'Hertogenbosh, the Netherlands (Dr J.W. Hoekstra); Inselspital, Bern, Switzerland (Prof U.E. Studer); Allgemeine Ziekenhuis Middelheim, Antwerpen, Belgium (Dr P. van Erps); Medical Radiological Research Center, Obninsk, Russia (Dr O. Kariakine); Ospedale di Modena S. Agostino Estense, Modena, Italy (Prof M. Brausi); Universitair Ziekenhuis Antwerpen, Antwerpen, Belgium (Dr L. Hoeckx); Cliniques Universitaires Saint Luc, Brussels, Belgium (Prof P. van Cangh); Univesita di Brescia, Brescia, Italy (Prof S. Magrini); Hospital Santa Maria, Lisboa, Portugal (Prof J.L. Carneiro di Moura); Trial 30903: Radium Hopsital, Oslo, Norway (Prof S.D. Fossa); Sint Antonius Ziekenhuis, Nieuwegein, the Netherlands (Dr P.H.T.J. Slee); Ospedale di Modena S. Agostino Estense, Modena, Italy (Prof M. Brausi); Antoni Vanleewenhookhuis, Amsterdam, the Netherlands (Dr S. Horenblas); Freeman Hospital, Newcastle-on-Tyne, United Kingdom (R.R. Hall); Princess Royal Hospital, Hull, United Kingdom (J. Hetherington); St James Hospital, Leeds, United Kingdom (P. Whelan); University School of Medicine, Izmir, Turkey (Prof Z. Kirkali); Erasmus University Medical Centrum, Rotterdam, the Netherlands (Dr J. Blom and Dr J. Klijn); Inselspital, Bern, Switzerland (Prof U. Studer); Medical Radiological Research Center, Obninsk, Russia (Dr O. Kariakine); Ospital Desterro, Amadorra, Portugal (Dr F. Calais da Silva); Allgemeine Ziekenhuis Middelheim, Antwerpen, Belgium (Dr P. Van Erps); Sint Radboud University Hospital, Nijmegen, the Netherlands (Prof F. Debruyne); Nemocnice Kromeriz, Kromeriz, Czech Republic (Dr L. Domes); Allgemeine Ziekenhuis Vrije University, Brussels, Belgium (Prof F. Keuppens); Allgemeine Ziekenhuis Vrije University, Amsterdam, the Netherlands (Prof D. Newling); Bosch Medicentrum, s'Hertogenbosh, the Netherlands (Dr J.W. Hoekstra); Ospedale Molinette, Torino, Italy (Prof A. Tizzani); Kantonspital, Aarau, Switzerland (Dr F. Recker); Comenius University Hospital, Martin, Slovakia (Prof J. Kliment); Trial 30944: Krankenanstalt Rudolfstiftung, Vienna, Austria (Dr W. Albrecht); Antoni Van Leeuwenhoekhuis, Amsterdam, the Netherlands (Dr S. Horenblas); Hopital Edouard Herriot, Lyon, France (Dr J.M. Maréchal); Erasmus Univesity Medical Centrum, Rotterdam, the Netherlands (Prof F.H. Schröder; Ospedale Molinette, Torino, Italy (Prof A. Tizzani); Royal Marsden Hospital, Surrey, United Kingdom (Prof A. Horwich); Universita Di Palermo, Palermo, Italy (Prof M. Pavone-Macaluso); Centro De Raferimento Oncologico, Aviano, Italy (Prof U. Tirelli); Comenius University Hospital, Martin, Slovakia (Prof J. Kliment); Freeman Hospital, Newcastle-on-Tyne, United Kingdom (R.R. Hall); Ospedale di Modena S. Agostino Estense, Modena, Italy (Prof M. Brausi); Ospedale di Circolo e Fundacione Macchi, Varese, Italy (Prof A. Bono); Hospices Civils de Strasbourg, Strasbourg, France (Prof D. Jacqmin); General Hospital, Weiner Neustadt, Austria (Dr F. Fouroutan); Hotel-Dieu De Paris, Paris, France (Dr S. Houdard); Marmara University Hospital, Istanbul, Turkey (Prof A. Akdas); Hopital Bichat-Claude Bernard, Paris (Prof L. Boccon-Gibod); Bosch Medicentrum, s'Hertogenbosh, the Netherlands (Dr J.W. Hoekstra); Institut Gustave Roussy, Villejuif, France (Dr C. Théodore); National Institute of Oncology, Budapest, Hungary (Dr I. Bodrogi); Onze Lieve Vrouw Ziekenhuis, Aalst, Belgium (Dr P. Carpentier); Universitair Ziekenhuis Gasthuisberg, Leuven, Belgium (Prof L. Baert); Universitair Ziekenhuis Gent, Gent, Belgium (Dr W. Oosterlinck); Donauspital Der Stadt Wien, Wien, Austria (Dr G. Studler); Kantonsspital, Basel, Switzerland (Dr R. Cassella); Princess Royal Hospital, Hull, United Kingdom (J. Hetherington); Sint Antonius, Nieuwegen, the Netherlands (Dr P.H.Th.J. Slee); and Centre Universitaire Hautepierre, Strasbourg, France (Prof J.P. Bergerat).

Authors' Disclosures of Potential Conflicts of Interest

The following authors or their immediate family members have indicated a financial interest. No conflict exists for drugs or devices used in a study if they are not being evaluated as part of the investigation. Consultant/Advisory Role: Sophie D. Fossa, Bayer. Research Funding: Sophie D. Fossa, Aventis. For a detailed description of these categories, or for more information about the American Society of Clinical Oncology's conflict of interest policy, please refer to the Author Disclosure Declaration form and the Disclosures of Potential Conflicts of Interest section of Information for Contributors found in the front of every issue.