treated with interferon alfa has been developed and validated through use of a large dataset. [J Natl Cancer Inst 1998;90:850 8]

Transcription

1 A New Prognostic Score for Survival of Patients With Chronic Myeloid Leukemia Treated With Interferon Alfa Joerg Hasford, Markus Pfirrmann, Rüdiger Hehlmann, Norman C. Allan, Michele Baccarani, Johanna C. Kluin-Nelemans, Guiliana Alimena, Juan Luis Steegmann, Hassan Ansari* Writing Committee for the Collaborative CML Prognostic Factors Project Group Background: Interferon alfa is a conservative and widely used alternative to bone marrow transplantation in treatment of patients with early chronic myeloid leukemia (CML). A meta-analysis was conducted to develop a reliable prognostic scoring system for estimation of survival of patients with CML treated with interferon alfa. Methods: Patients treated in prospective studies, including major randomized trials, were separated into learning and validation samples. Cox regression analysis and the minimum P-value approach were used to identify prognostic factors for patient survival and to discover groups in the learning sample with the greatest differences in survival. These findings were then validated by applying the new scoring system to patients in the validation sample. Results: We collected data on 1573 patients who were participants in 14 studies involving 12 institutions; 1303 patients (learning sample, n = 981; validation sample, n = 322) were eligible for inclusion in this analysis, and their median survival time was 69 months (range, months). Because two previously described prognostic scoring systems failed to discriminate risk groups satisfactorily, we developed a new scoring system that utilizes the following covariates: age, spleen size, blast count, platelet count, eosinophil count, and basophil count. Among 908 patients with complete data in the learning sample, three distinct risk groups were identified (median survival times of 98 months [n = 369; 40.6%], 65 months [n = 406; 44.7%], or 42 months [n = 133;14.6%]; two-sided logrank test, P.0001). The ability of the new scoring system to discriminate these risk groups was confirmed by analysis of 285 patients with complete data in the validation sample (two-sided logrank test, P =.0002). Conclusions: A new prognostic scoring system for estimating survival of patients with CML treated with interferon alfa has been developed and validated through use of a large dataset. [J Natl Cancer Inst 1998;90:850 8] Over the last decade, interferon alfa (IFN ) became accepted as the best available conservative treatment for patients with early stage chronic myeloid leukemia (CML). Allogeneic bone marrow transplantation, regarded as the only curative therapy, still carries a considerable risk of early death. Hence, there is a need to assess the prognosis of patients treated with IFN. The Sokal score (1) works well as a prognostic discriminator for survival of three different strata of patients treated with busulfan or hydroxyurea (i.e., groups with low risk, intermediate risk, or high risk of death), but it has repeatedly been reported to be a poor discriminator for survival of patients treated with IFN (2 4). Kantarjian et al. (5) developed a new staging system for the prediction of survival of patients with distinct clinical features. Examined in a cohort of German patients by Hasford et al. (3), this staging system was also not able to distinguish patient groups with different survival. Prognostic factors serve a variety of important tasks in modern medicine (6): selection of the optimal treatment, development of risk-adjusted treatments, adjustment of imbalances of treatment groups in clinical trials, and comparative assessment of the outcomes of different studies. For these reasons, the Collaborative CML Prognostic Factors Project was started. The goal of this project was to extract and validate a new prognostic scoring system for patients with CML treated with IFN. Methods Data Collection In addition to established contacts with other CML study groups, we identified studies relevant to our study objectives by searching the MEDLINE biomedical literature database (National Library of Medicine, Bethesda, MD), by screening abstracts and conference reports, and by contacting pharmaceutical companies. All datasets had to be from prospective studies of patients with CML. See the Appendix section for the address of the center for data collection and analysis as well as names and addresses of further members of the Collaborative CML Prognostic Factors Project. We received data on 1573 individual patients from 14 studies (see Appendix Table 1) involving 12 centers. Ten investigators sent their data by electronic files (2,7 17); the data of four studies [(18 20) and the data submitted by L. Dabich] were sent by use of our standard case report form. Participants were part of studies conducted in Austria (7,8); Belgium, The Netherlands, and Luxembourg (9,10); France (11,12); Germany (13,14,18); the U.K. (15); Italy (16,19); Japan (2); Spain (17,20); and the United States. So far, the U.S. patients (L. Dabich) have not been part of a publication. A few relevant details regarding each study are included in Appendix Table 1. Characteristics of Patients All important clinical features of the participants are described in Table 1. Bone marrow features regarding megakaryocytes, basophils, and pseudo- Gaucher cells as well as fibrosis were not included, since hardly any of that information on participants was available. Investigators were asked whether the Philadelphia chromosome had been detected (if so, *Affiliations of Writing Committee: J. Hasford, M. Pfirrmann, Institute for Medical Informatics, Biometry, and Epidemiology, University of Munich, Germany; R. Hehlmann, III. Medizinische Klinik, Universität Heidelberg, Mannheim, Germany; N. C. Allan, Medical Research Council Human Genetics Unit, Western General Hospital, Edinburgh, U.K.; M. Baccarani, Cattedra di Ematologia, Università di Udine, Italy; J. C. Kluin-Nelemans, Department of Hematology, Leiden University Medical Center, The Netherlands; G. Alimena, Dipartimento di Biopatologia Umana, Sezione Ematologia, Universita degli Studi La Sapienza, Rome, Italy; J. L. Steegmann, Hospital Universitario de la Princesa, Servicio de Hematología, Madrid, Spain; H. Ansari, Biometric Center for Therapeutic Studies, Munich, Germany. Correspondence to: Joerg Hasford, M.D., Ph.D., Institute for Medical Informatics, Biometry and Epidemiology (IBE), University of Munich, Marchioninistr. 15, Munich, Germany. has@ibe.med.uni-muenchen.de See Notes following References. Oxford University Press 850 REPORTS Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998

2 Baseline variable (No. of patients) the patient was called Ph-positive) and to report whether additional chromosome abnormalities, other than the chromosome 9:chromosome 22 translocation (bcr/abl), had been found. We required that all data regarding the clinical features had to be recorded at the time of the patients diagnosis. Important items included in this analysis concerning the treatment of the patients in all studies were date of diagnosis, kind of pretreatment, allocated and actual study treatment, dates of starting and stopping treatment, and reason for stopping treatment. Admission Criteria Table 1. Initial characteristics of most important baseline variables Mean (CI*) or proportion Median Range Age, y (1303) 47 (28 75) Sex, proportion male (1303) 0.58 Hemoglobin, g/dl (1275) 11.9 ( ) White blood cell count, 10 9 /L (1291) 143 (0 349) Platelet count, 10 9 /L (1292) 505 (0 1199) Blasts, % (1258) 1.8 (0 6.1) Promyelocytes, % (1166) 3.5 (0 11.3) Basophils, % (1283) 4.1 (0 11.2) Eosinophils, % (1279) 2.6 (0 7.5) Erythroblasts, (767) 1.1 (0 8.5) Blasts in bone marrow, % (734) 2.5 (0 7.4) Promyelocytes in bone marrow, % (496) 7.5 (0 20.4) Additional chromosome abnormalities, proportion (861) 0.08 Fatigue, general ill feeling, proportion (985) 0.48 Symptoms due to organomegaly, proportion (916) 0.22 Weight loss of more than 10%, proportion (844) 0.18 Fever, proportion (1016) 0.07 Karnofsky index (783) 94 (76 100) Smoker, proportion (252) 0.19 Spleen size, cm below costal margin (1240) 5.4 (0 17.2) Hepatomegaly, proportion (1089) 0.33 Time from diagnosis to start of treatment, days (1303) 44 (0 136) Follow-up of all patients, days (1303) 1322 (0 2912) Follow-up of patients censored alive, days (800) 1340 (0 3020) Follow-up only of patients with bone marrow transplantation in first chronic phase (228) 767 (0 1955) *Standard 95% confidence intervals. Apart from the Philadelphia chromosome 9: chromosome 22 translocation and the abnormalities that led to exclusion. Reference (21). Patients had to have been treated with IFN, either alone or in combination with another therapy (hydroxyurea, cytarabine, or IFN gamma). Further inclusion criteria were a confirmed diagnosis of Phpositive or bcr/abl-positive CML, known date of diagnosis, and details of any pretreatment. A patient was excluded from the study when any of the following criteria was violated: minimum observation time of at least 1 day, initiation of treatment not more than half a year after diagnosis, initiation of treatment during the period from January 1, 1983, through December 31, 1994, or patient s disease in chronic phase on the date of diagnosis. To ensure that the disease in all patients was in the chronic phase, we applied the criteria established by the Italian Cooperative Study Group on Chronic Myeloid Leukemia (16) and removed patients who displayed any of the following features: greater than 10% blasts in the peripheral blood, greater than 30% (blasts + promyelocytes) in the peripheral blood, greater than 15% blasts in bone marrow, greater than 50% (blasts + promyelocytes) in bone marrow, extramedullary manifestations, two Philadelphia chromosomes, trisomy 8, or isochromosome 17. Patients for whom one or more of the Italian criteria could not be examined because of missing values were not excluded. First, of the examined patients, only a small percentage (6.2%) actually had to be excluded. Second, there was no difference in survival between the patients for whom all criteria could have been checked and whose disease was definitely not in the accelerated/blastic phase and the patients for whom at least one criteria could not be examined but whose disease was not in accelerated/blastic phase according to all other criteria. Finally, excluding patients for whom not all Italian criteria could be examined would have reduced the final sample by several hundred patients. Data Checking Various checks for logical sequence and consistency of the data were performed. Each study was analyzed separately. We checked the data for outliers and categorial variables for unexplained values. Calendar dates were examined for their natural order and for completeness. Contents of directly dependent items and the reasons for missing values were closely scrutinized. Statistical Analysis After all inclusion and exclusion criteria were applied, and before any analyses were conducted, patients were divided into a learning sample and a validation sample. It was our goal to have a learning sample that consisted of approximately 75% of all patients eligible for analysis. At first, one (11,12) of the four studies (11,12,14 16) containing more than 100 patients was randomly drawn to become part of the validation sample. Additional studies (13,19) were randomly drawn from the remaining pool of 13 studies (2,7 10,13 20) until a validation sample that included 20% 30% of all eligible patients was obtained. For all subsequent analyses, chi-squared tests, Student s t tests, logrank tests, Kaplan Meier survival analyses, Cox regression analyses, and minimal P-value approach tests were applied with use of SAS procedures and programs (SAS Institute, Inc., Cary, NC) (22). All P values resulted from the use of two-sided statistical tests. Survival Time Survival time was measured as the interval between the start of study treatment and death or last follow-up date. Patients who received a bone marrow transplant at the time when their disease was in its first chronic phase were censored on the date of the transplant procedure. For patients whose disease had already been in second chronic phase, blastic phase, or accelerated phase when undergoing a bone marrow transplantation, the event of bone marrow transplantation was not censored since the preceding IFN treatment had already failed. Analysis of the Learning Sample To identify discrete covariates that were univariately significant for survival, we applied logrank tests and plotted the corresponding Kaplan Meier curves. Univariate associations between continuous variables and survival were evaluated by Cox regression analyses. A statistical significance level of was chosen. The proportional hazards assumption was examined by comparison of grouped survival curves and the analysis of residuals (23), respectively. The minimum P-value approach was used for categorization of continuous variables (24 28). All statistically significant prognostic factors from a multivariate Cox model were selected by stepwise regression as described for the corresponding SAS procedure [(22), pages 832, 842 8]: At a certain step, a variable is entered into the model if the value of its likelihood score test for testing the significance of the variable in the presence of the previously entered variables is significant (significance level 2 0.1) and also higher than the likelihood score values of all other variables in question. A variable is removed if its Wald statistic is not significant (significance level 2 0.1). Only patients with complete data with regard to all factors of a certain multivariate model were included for the analysis of that model. Confirmation was sought by application of a strategy for model selection suggested by Collett [(29), p. 80 5]. To determine correlations between covariates, scatterplots were drawn, Pearson correlation coefficients were measured, and chi-squared tests and t tests were performed. We checked whether interactions of multivariately significant variables could improve the model, and also we reintroduced univariately unimportant variables to examine their statistical significance in the presence of other variables. For the investigation of alternative models, highly correlated covariates were exchanged. The prognostic score of the individual patients was calculated by combining the selected prognostic factors gained Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998 REPORTS 851

3 from the learning sample. With the use of these data, the minimum P-value approach helped to determine the statistically significant cut points that divided the patients of the learning sample into corresponding risk groups. Analysis of the Validation Sample The new prognostic score and the cut points evaluated in the learning sample were then applied to the validation sample. To investigate whether a certain risk group of the validation sample would generate a survival curve that would be distinguishable in a statistically significant way from the survival curves of all differing risk groups of the learning sample, we performed pairwise logrank tests comparing each risk group of the validation sample with each differing risk group of the learning sample. Results The final dataset consisted of 1303 patients who fulfilled the admission criteria. Their median age was 49 years (range, years); 58% of the patients were men (Table 1). Their median survival time was 69 months (range, months). Information on each of the variables, i.e., hemoglobin, white blood cell count, blast count, eosinophil count, basophil count, platelet count, or spleen size, was available for more than 95% (n 1238) of the case subjects. The median follow-up time was 39 months for all patients and 40 months for the 800 living (censored) patients (both ranges, 12 days 124 months). Two hundred twenty-eight of 800 patients were censored because of allogeneic bone marrow transplantation in the first chronic phase of their disease. Of the 503 patients who died, the cause of death was assessed to be CML related in 404 case subjects, not CML related in 68 case subjects, and unknown in 31 case subjects. Examination of Scores Already Published When the Sokal score was applied to 1201 assessable patients (Fig. 1, A), it did not distinguish intermediate-risk patients from high-risk patients. Their survival curves separated only after 2.5 years and met again after 6.25 years. The difference in median survival between these two groups was less than 1 year. When applying the second Sokal score (30), devised for patients less than 46 years old at time of diagnosis and with age replaced by the prognostic factors of sex and hematocrit, results were not better. When the new staging system of Kantarjian et al. (5) was used (Fig. 1, B), the survival curve of the patients with the second best of four stages was below all other curves. None of the four survival curves displayed a clear separation from all other survival curves, and survival curves cross each other several times. Split in Learning and Validation Sample Three studies (11 13,19), altogether with 322 patients (24.7% of the eligible patients), were drawn for the validation sample. Hence, the learning sample (2,7 10,14 18,20) consisted of 981 patients (75.3%). There was no statistically significant difference (significance level ) between the two samples with regard to survival or censoring pattern. Statistically significant differences for the covariate spleen size (cm below left costal margin, P.010), hemoglobin (g/dl, P.017), bone marrow promyelocytes (%, P.001), bone marrow fibrosis (P.015), and hepatomegaly (P.016) indicate the independence of the two patient samples. Changing the Cut Points of the Sokal Score in the Learning Samples At first, we calculated the Sokal score for all 912 of the assessable patients who had been included in the learning sample. Applying the minimal P-value approach, we tried to improve the original cut points (0.8 and 1.2) of Sokal et al. (1). We obtained cut points 0.604, 0.955, and instead, but the performance in the validation sample was poor: The survival curve of the high-risk group crossed the survival curves of the low- and the intermediate-risk groups several times. Thus, the Sokal score could not be improved, and a new score had to be developed. Analysis of the Learning Sample The clinical features (Table 1) showing a univariately significant correlation with survival were age, sex, spleen size (cm below left costal margin), white blood cell count ( 10 9 /L), blasts (%), promyelocytes (%), basophils (%), eosinophils (%), erythroblasts (%), the preceding five features are all given as proportions in white blood cells in the peripheral blood, hemoglobin (g/dl), blasts (%) in bone marrow, weight loss, and hepatomegaly (Table 2). Proportional hazards assumptions were met. Erythroblasts, weight loss, and hepatomegaly were not statistically significant in a multivariate model, but blasts in bone marrow remained significant. Age contributed to a better fit when used as a dichotomous variable (age 0 if age <50 years; age 1, otherwise). Only as dichotomized factors did basophils (basophils 0 if basophils < 3%; basophils 1, otherwise) and platelets (platelets 0 if platelets < /L; platelets 1, otherwise) reach statistical significance in a multivariate model. Further variates that were dichotomized according to a cut point identified by the minimal P-value approach were spleen size (cut point 7 cm), white blood cell count ( /L), blasts (1%), and hemoglobin (11 g/dl), none of which could improve our final model. The final model consisted of the variates of age, spleen size, blasts, basophils, platelets, and eosinophils that were selected by the stepwise procedure as well as by Collett s strategy (29) for model selection (significance level 2 0.1, Table 2). Age, basophils, and platelets were entered into the model as dichotomized factors. We did not consider blasts in bone marrow because this information was available only in 56% of all case subjects. No further variate, dichotomized factor, interaction, or mixed term, formed from a factor and a variate, was statistically significant when added to our final model. The coefficients of the final model were evaluated from the 908 complete case subjects when only the six selected variates were included. New Prognostic Score We obtained the following new prognostic score: New prognostic score ( age [0 when age <50 years; 1, otherwise] spleen size [cm below costal margin] blasts [%] eosinophils [%] basophils [0 when basophils <3%; 1, otherwise] platelet count [0 when platelets < /L; 1, otherwise]) The median score value of the above 908 patients was 913, the lowest value was 0, the highest value was 2697, and the mean value was 916. Using the minimum P-value approach, we identified three groups who differed significantly with regard to survival times. Patients in the low-risk group had score values 852 REPORTS Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998

4 Fig. 1. A) The Sokal score (1) applied to 1201 patients with early stage chronic myeloid leukemia treated with interferon alfa (learning and validation sample). Low risk: Sokal score <0.8; intermediate risk: 0.8 Sokal score 1.2; high risk: Sokal score >1.2. Entry (425/110), 96 months indicates that, of 425 patients, 110 patients died. Median survival time was 96 months (range, months). The same meaning applies to the numbers of the intermediate-risk group (range, months) and the high-risk group (range, months). Lengths of horizontal crossbars indicate the upper and lower 95% confidence limits. Crossbars increase from the lowest risk group to the highest risk group. B) The staging system of Kantarjian et al. (5) was applied to 660 patients with early stage chronic myeloid leukemia treated with interferon alfa (learning and validation sample). Entry (393/104), 79 months indicates that of 393 patients, 104 patients died. Median survival time was 79 months (range, 1 96 months). The same meaning applies to the numbers of stage 2 (range, 2 88 months), stage 3 (range, months), and stage 4 (range, 6 85 months). Lengths of horizontal crossbars indicate the upper and lower 95% confidence limits. Crossbars increase from the lowest risk group to the highest risk group. 780, the intermediate-risk group had score values >780 and 1480, and the high-risk group had score values >1480. The 369 patients of the low-risk group (40.6% of all patients) had a median survival time of 98 months (range, months) (Fig. 2, A), the intermediate-risk group (n 406, 44.7% of all patients) of 65 months (range, months), and the high-risk group (n 133, 14.6% of all patients) of 42 months (range, months). The 5-year survival rate was 76% (95% confidence interval [CI] 70% 82%) in the low-risk group, 55% (95% CI 48.6% 61.4%) in the inter- Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998 REPORTS 853

5 Table 2. Significant prognostic factors, P values from Wald s chi-squared statistic (two-sided) Clinical feature mediate-risk group, and 25% (95% CI 15% 35%) in the high-risk group. Validation of the New Prognostic Scoring System When the new scoring system was applied to the 285 assessable patients from the three independent studies of the validation sample, 125 case subjects (43.8%) were assigned to the low-risk group, 125 (43.8%) to the intermediate-risk group, and 35 (12.2%) to the high-risk group. Relative sample sizes and median survival times of the three risk groups (96, 65, and 45 months, respectively) were hardly any different from the results obtained for the corresponding risk group in the learning sample (Fig. 2, B). The 5- year survival rates of the low-risk group (73%; 95% CI 63.4% 82.6%) and of the intermediate-risk group (55%; 95% CI 45% 65%) were similar to the 5-year survival rates in the learning sample. The 5-year survival rate of the high-risk group was 37% (95% CI 19% 55%). The pairwise comparison of each risk group in the validation sample to each differing risk group in the learning sample always led to statistically significant results (significance level ). Various Features of the New Prognostic Scoring System Univariate analysis (2 0.1) Multivariate analysis: final model (2 0.1), 908 case subjects Age (0/1)* P.0001 P.0001, in final model Spleen size P.0001 P.0001, in final model Hemoglobin P case subjects, P.6494 White blood cells P case subjects, P.8864 Blasts P.0345 P.0214, in final model Promyelocytes P case subjects, P.8845 Eosinophils P.0035 P.0868, in final model Basophils (0/1) P.0017 P.0835, in final model Platelet count (0/1) P.0001 P.0002, in final model Bone marrow blasts P case subjects, P.0877 Weight loss P case subjects, P.1851 Hepatomegaly P case subjects, P.8450 Erythroblasts P case subjects, P.8968 *Age (0/1): age 0 if age <50 years; age 1, otherwise. Number of case subjects when factor was added to the six statistically significant factors of the final model. P values are evaluated from the model including this factor and the six statistically significant factors of the final model. Basophils (0/1): basophils 0 if basophils <3%; basophils 1, otherwise. Platelet count (0/1): platelet count 0 if platelet count < /L; platelet count 1, otherwise. We stratified the 1193 assessable patients by country and obtained four large study groups for France (169 patients), Germany (255 patients), Italy (225 patients), and the U.K. (186 patients). All countries providing fewer than 100 patients were grouped together (358 patients). The new scoring system clearly separated the survival curves of each of the five strata statistically significantly (two-sided logrank test: France, P.0027; Germany, P.0339; Italy, P.0001; U.K., P.0002; and other countries, P.0001). We further applied the Sokal score and our new scoring system to 1021 chemotherapy patients. The Sokal score allocated 301 patients (29.5%) to the low-risk group (our scoring system: 340 patients, 33.3%), 352 patients (34.5%) to the intermediate-risk group (our scoring system: 482 patients, 47.2%), and 368 patients (36%) to the high-risk group (our scoring system: 199 patients, 19.5%). Median survival times of the three risk groups were approximately the same for both scores: For the low-risk group, they were 62 months (Sokal score; range, months) and 60 months (new scoring system; range, months); for the intermediate-risk group, they were 52 months (Sokal score; range, months) and 49 months (new scoring system; range, months); and for the high-risk group, they were 39 months (Sokal score; range, months) and 38 months (new scoring system; range, 2 93 months). While the three survival curves obtained from our scoring system clearly separated right from the beginning, the survival curves of the low- and the intermediate-risk groups of the Sokal score did not separate for the first 3.5 years. Because we did not think a strict intention-to-treat analysis would provide us with the proper patient sample in which we should perform our prognostic factors analysis, we had excluded 270 patients for the reasons outlined above. However, we examined the performance of the new scoring system in 248 of those patients for whom the new score was assessable. The low-risk group consisted of 85 patients (34.3%), the intermediate-risk group was made up of 103 patients (41.5%), and the high-risk group consisted of 60 patients (24.2%). Median survival times were 97 months (range, months), 67 months (range, 4 days to 103 months), and 32 months (range, 5 82 months), respectively. That the median survival of the high-risk group was 10 months less than in the learning sample and that the percentage of patients in the high-risk group amounted to 24.2% rather than 14.6% (learning sample) are not surprising results, since 36.7% of the 270 patients were excluded because their disease was already in the accelerated or blastic phase. Discussion The major objective of this metaanalysis was to extract and validate a prognostic scoring system for patients with chronic-phase CML treated with IFN. The need for a new prognostic scoring system was reinforced by the finding that the scoring systems of Sokal et al. (1,30) and of Kantarjian et al. (5) do not separate survival curves well. We succeeded in extracting from the data a new scoring system for patients treated with IFN that distinguishes three risk groups, both in learning and in validation samples. Merging the two samples, the new scoring system displays a low-risk group with a median survival of 96 months (range, months) and a 5-year survival rate of 75% (95% CI 70% 80%). The low-risk group represents 41.4% of all 1193 assessable patients. The high-risk group contained 168 patients with a median survival time of 42 months (range, months); these patients had a 5-year survival rate of 28% (95% CI 19% 37%) and represented 14.1% of the whole sample. The median survival time of pa- 854 REPORTS Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998

6 Fig. 2. A) The new prognostic scoring system applied to 908 patients with early stage chronic myeloid leukemia treated with interferon alfa (learning sample). Low risk: new score 780; intermediate risk: 780 < new score 1480; high risk: new score >1480. Entry (369/83), 98 months indicates that of 369 patients, 83 patients died. Median survival time was 98 months (range, months). The same meaning applies to the numbers of the intermediate-risk group (range, months) and the high-risk group (range, months); two-sided logrank test: P Lengths of horizontal crossbars indicate the upper and lower 95% confidence limits. Crossbars increase from the lowest risk group to the highest risk group. B) The new prognostic scoring system applied to 285 patients with early stage chronic myeloid leukemia treated with interferon alfa (validation sample). Low risk: new score 780; intermediate risk: 780 < new score 1480; high risk: new score >1480. Because few patients were observed for longer, only the survival curves for the first 5.5 years under observation were plotted. Entry (125/33), 96 months indicates that of 125 patients, 33 patients died. Median survival time was 96 months (range, months). The same meaning applies to the numbers of the intermediate-risk group (range, 7 84 months) and the high-risk group (range, 2 88 months); two-sided logrank test: P Lengths of horizontal crossbars indicate the upper and lower 95% confidence limits. Crossbars increase from the lowest risk group to the highest risk group. tients in the low-risk group was more than twice as long as that in the high-risk group. The intermediate-risk group made up 44.5% of the assessable patients; these patients had a median survival of 65 months (range, months) and a 5- year survival rate of 56% (95% CI 51% 61%). Almost all prospective studies of CML treated with IFN that have been conducted worldwide are part of our database. There was no difference in survival Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998 REPORTS 855

7 Appendix Table 1. All studies that were included Study* Year study started No. of patients in final sample No. of patients who died No. of patients censored Median months of follow-up of survivors Austrian CML III (7) Austrian CML V (7,8) Benelux Study Group (9,10) Castilla-León (20) Dabich et al German CML II (14) Italian CML 86 (16) Kloke et al. (18) Kouseisho Study Group (2) MRC CML III (15) Spanish Study Group (17) Alimena et al. (19) French CML 88 (11,12) German CML I, IFN (13) *The last three studies belonged to the validation sample. CML chronic myeloid leukemia; MRC Medical Research Council; IFN interferon. Results not published, but study protocol was available. between the various IFN treatments, whether IFN was administered alone or as part of the drug combinations mentioned above (P =.8056). The only studies included in our meta-analysis that compared IFN treatment with a combination therapy involving IFN and cytarabine (8,12) also have not yet found any significantly different survival curves. The data of the French CML 91 trial, for which results have recently been published (31) were not part of our meta-analysis. The median survival of our final set of 1303 patients was 69 months, 7 months less than the median survival of the Italian patients with CML and 3 months less than the median survival of the German patients with CML when a comparable definition of chronicphase disease was applied (32). The patients included in the analysis conducted by Kantarjian et al. (33) had a median survival of 89 months. The lower median survival time and the median age of 49 years suggest that our patient sample is representative and not a subselection of patients with a relatively good prognosis. A few patients might have had disease in the accelerated/blastic phase, as the necessary data to check for accelerated/ blastic phase were missing, but this situation should not have a considerable impact on the new scoring system. The variables of the new scoring system are routinely measured in clinical practice, and their measurement is highly reliable, with the probable exception of spleen size. Thus, there should not be a problem to get reliable data on the variables needed for the calculation of the new prognostic score. The new scoring system has shown a good discrimination of survival in a representative patient sample and has also been thoroughly validated. A variety of considerations underline its importance. The new scoring system allows a better planning and comparative interpretation of the results of different trials, because differing prognostic profiles of patient samples can override treatment effects and thus bias results (32,34). We hope that the new scoring system will also be considered for the development and the evaluation of risk-adapted treatments. Appendix Writing committee Professor Dr. Joerg Hasford, Institute for Medical Informatics, Biometry and Epidemiology (IBE), University of Munich, Marchioninistr. 15, Munich, Germany. has@ibe.med.uni-muenchen.de Markus Pfirrmann, M.Sc., Institute for Medical Informatics, Biometry and Epidemiology (IBE), University of Munich, Marchioninistr. 15, Munich, Germany. pfi@ibe.med.uni-muenchen.de Professor Dr. Rüdiger Hehlmann, III. Medizinische Klinik, Universität Heidelberg, Wiesbadener Str. 7-11, Mannheim, Germany. Professor Dr. Norman C. Allan, MRC Human Genetics Unit, Western General Hospital, Crewe Road, Edinburgh EH4 2XU, U.K. Professor Dr. Michele Baccarani, Cattedra di Ematologia, Università di Udine, P. le S.M. della Misericordia, Udine, Italy. Dr. J. C. Kluin-Nelemans, Department of Hematology, Leiden University Medical Center, P.O. Box 9600, 2300 RC Leiden, The Netherlands. Professor Dr. Giuliana Alimena, Dipartimento di Biopatologia Umana, Sezione Ematologia, Universita degli Studi La Sapienza, Via Benevento, 6, Roma, Italy. Dr. Juan Luis Steegmann, Hospital Universitario de la Princesa, Servicio de Hematología, C/. Diego de León, Madrid, Spain. Dr. Hassan Ansari, Biometric Center for Therapeutic Studies (BZT), Pettenkoferstr. 35, Munich, Germany. Center for data collection and analysis Institute for Medical Informatics, Biometry and Epidemiology (IBE), University of Munich, Marchioninistr. 15, Munich, Germany. has@ibe.med.uni-muenchen.de Further members of the Collaborative CML Prognostic Factors Project The Italian Cooperative Study Group on Chronic Myeloid Leukemia was represented by Prof. Dr. Sante Tura (Chairman) and Dr. Eliana Zuffa (Secretary General), Istituto di Ematologia L.e.A. Seraguoli, Policlinico S. Orsola, via Massarenti 9, 40138, Bologna, Italy. Professor Dr. L. Dabich, Department of Internal Medicine, Division of Hematology/ Oncology, The University of Michigan, 5301 MSRB III, 1150 W. Medical Center Drive, Ann Arbor, MI The Benelux CML Study Group was further represented by Dr. A. Delannoy, Hôpital de Jolimont, 7100 Haine-St-Paul, Belgium and Dr. A. Louwagie, Department of Hematology, AZ St. Jan Ruddershove 10, 8000 Brugge, Belgium. The French CML Study Group was represented by Professor Dr. François Guilhot, Départment d Hématologie et Oncologie médicale, Hôspital Jean Bernard, 350 Avenue Jacques Coeur B.P. 577, Poitiers, France. 856 REPORTS Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998

8 Privatdozent Dr. O. Kloke, Abteilung Innere Medizin/Krebszentrum, Universitätsklinikum Essen, Hufelandstr. 55, Essen, Germany. Dr. M. J. Moro, Castilla León Cooperative Group, Hospital de León, Virgen Blanca, León, Spain. The Kouseisho Leukemia Study Group was represented by Professor Dr. Kazunori Ohnishi, and Professor Dr. R. Ohno, Internal Medicine III, Hamamatsu University, School of Medicine, 3600 Handa-cho, Hamamatsu , Japan. The U.K. Medical Research Council s Working Parties for Therapeutic Trials in Adult Leukemia was further represented by Dr. S. Richards, Clinical Trial Service Unit, Harkness Building/Radcliffe Infirmary, Oxford OX2 6HE, U.K. The Austrian CML Study Group was represented by Professor Dr. Josef Thaler, Universitatsklinik für Innere Medizin, Anichstraße 35, 6020 Innsbruck, Austria. References (1) Sokal JE, Cox EB, Baccarani M, Tura S, Gomez GA, Robertson JE, et al. Prognostic discrimination in good-risk chronic granulocytic leukemia. Blood 1984;63: (2) Ohnishi K, Ohno R, Tomonaga M, Kamada N, Onozawa K, Kuramoto A, et al. A randomized trial comparing interferon-alpha with busulfan for newly diagnosed chronic myelogenous leukemia in chronic phase. Blood 1995;86: (3) Hasford J, Ansari H, Pfirrmann M, Hehlmann R. Analysis and validation of prognostic factors for CML. German CML Study Group. Bone Marrow Transplant 1996;17(Suppl 3):S49 S54. (4) Ozer H, George SL, Schiffer CA, Rao K, Rao PN, Wurster-Hill DH, et al. Prolonged subcutaneous administration of recombinant alpha 2b interferon in patients with previously untreated Philadelphia chromosome-positive chronic-phase chronic myelogenous leukemia. Effect on remission duration and survival: Cancer and Leukemia Group B Study Blood 1993;82: (5) Kantarjian HM, Keating MJ, Smith TL, Talpaz M, McCredie KB. Proposal for a simple synthesis prognostic staging system in chronic myelogenous leukemia. Am J Med 1990;88: 1 8. (6) Byar DP. Identification of prognostic factors. In: Buyse ME, Staquet MJ, Sylvester RS, editors. Cancer clinical trials. Methods and practice. Oxford, England: Oxford University Press, 1984: (7) Thaler J, Gastl G, Fluckinger T, Niederwieser D, Huber H, Seewann H, et al. Treatment of chronic myelogenous leukemia with interferon alfa-2c: response rate and toxicity in a phase II multicenter study. The Austrian Biological Response Modifier (BRM) Study Group. Semin Hematol 1993;30(Suppl 3): (8) Thaler J and Hilbe W. Comparative analysis of two consecutive phase II studies with IFNalpha and IFN-alpha + ara-c in untreated chronic-phase CML patients. Austrian CML Study Group. Bone Marrow Transplant 1996;17(Suppl 3):S25 S28. (9) Department of Hematology, Leiden University Hospital, The Netherlands. Low-dose interferon alpha 2b combined with hydroxyurea versus hydroxyurea alone for chronic myelogenous leukemia. The Benelux CML Study Group. Bone Marrow Transplant 1996;17(Suppl 3):S19 S20. (10) Randomized study on hydroxyurea alone versus hydroxyurea combined with low-dose interferon-alfa2b for chronic myeloloid leukemia. Blood 1998;91: (11) Guilhot F. Interferon alfa and low-dose cytosine-arabinoside for the treatment of patients with chronic myelogenous leukemia in chronic phase. French CML Study Group. Semin Hematol 1993;30(Suppl 3):24 5. (12) Guilhot F, Guerci A, Fiere D, Harousseau JL, Maloisel F, Bouabdallah R, et al. The treatment of chronic myelogenous leukemia by interferon and cytosine-arabinoside: rationale and design of the French trials. French CML Study Group. Bone Marrow Transplant 1996;17: S29 S32. (13) Hehlmann R, Heimpel H, Hasford J, Kolb HJ, Pralle H, Hossfeld DK, et al. Randomized comparison of interferon-alpha with busulfan and hydroxyurea in chronic myelogenous leukemia. German CML Study Group. Blood 1994;84: (14) Hehlmann R, Heimpel H, Hossfeld DK, Hasford J, Kolb HJ, Loffler H, et al. Randomized study of the combination of hydroxyurea and interferon alpha versus hydroxyurea monotherapy during the chronic phase of chronic myelogenous leukemia (CML Study II) The German Study Group. Bone Marrow Transplant 1996;17(Suppl 3): S21 S24. (15) Allan NC, Richards SM, Shepherd PC. UK Medical Research Council randomised, multicentre trial of interferon-alpha n1 for chronic myeloid leukaemia: improved survival irrespective of cytogenetic response. The UK Medical Research Council s Working Parties for Therapeutic Trials in Adult Leukaemia. Lancet 1995;345: (16) Interferon-alpha-2a as compared with conventional chemotherapy for the treatment of chronic myeloid leukemia. The Italian Cooperative Study Group on Chronic Myeloid Leukemia. N Engl J Med 1994;330: (17) Fernandez-Ranada JM, Lavilla E, Odriozola J, Garcia-Larana J, Lozano M, Parody R, et al. Interferon alfa 2A in the treatment of chronic myelogenous leukemia in chronic phase. Results of the Spanish Group. Leuk Lymphoma 1993;11(Suppl 1): (18) Kloke O, Niederle N, Qiu JY, Wandl U, Moritz T, Nagel-Hiemke M, et al. Impact of interferon alpha-induced cytogenetic improvement on survival in chronic myelogenous leukaemia. Br J Haematol 1993;83: (19) Alimena G, Morra E, Lazzarino M, Liberati AM, Montefusco E, Inveradi D, et al. Interferon alpha-2b as therapy for Ph -positive chronic myelogenous leukemia: A study of 82 patients treated with intermittent or daily administration. Blood 1988;72: (20) Moro MJ, Gil S, Canizo C, Clemente JF, Guerras L, Fisac RM, et al. The treatment of chronic myelogenous leukemia with interferon alfa-2b plus hydroxyurea versus hydroxyurea alone. Haematologica 1991;76(Suppl 4): 117. (21) Karnofsky DA, Burchenal JH. The clinical evaluation of chemotherapeutic agents in cancer. In: Macleod CM, editor. Evaluation of chemotherapeutic agents. New York: Columbia University Press, (22) SAS Institute Inc. SAS/STAT Software: Changes and enhancements through Release Cary (NC): SAS Institute, Inc., (23) Harris EK, Albert A. Survivorship analysis for clinical studies. New York: Marcel Dekker, (24) Breiman L, Friedman JH, Olshen RA, Stone CJ. Classification and regression trees (CART). Belmont, CA: Wandsworth, (25) Brandmaier R, Ansari H, Dickson D. A general SAS macro for construction of classification and regression trees. In: Proceedings of the 18th annual SAS User Group International (SUGI) Conference, Cary (NC): SAS Institute, Inc., USA. 1993; (26) Ansari H, Aydemir Ü, Dirschedl P, Hasford J, Hehlmann R. Analyse von prognostischen Faktoren bei Patienten mit chronisch-myeloischer Leukámie (CML)-Ein Vergleich von CART- Technik und Cox-Modell. In: Michaelis J, Hommel G, Wellek S, editors. Europáische Perspektiven der Medizinischen Informatik, Biometrie und Epidemiologie. Medizinische Informatik, Biometrie und Epidemiologie 76. München:MMV Medizin Verlag, 1993: (27) Lausen B, Schumacher M. Maximally selected rank statistics. Biometrics 1992;48: (28) Altman DG, Lausen B, Sauerbrei W, Schumacher M. Dangers of using optimal cutpoints in the evaluation of prognostic factors. J Natl Cancer Inst 1994;86: (29) Collett D. Modelling survival data in medical research. London: Chapman & Hall, (30) Sokal JE, Baccarani M, Tura S, Fiacchini M, Cervantes F, Rozman C, et al. Prognostic discrimination among younger patients with chronic granulocytic leukemia: relevance to bone marrow transplantation. Blood 1985;66: (31) Guilhot F, Chastang C, Michallet M, Guerci A, Harousseau JL, Maloisel F, et al. Interferon alfa-2b combined with cytarabine versus interferon alone in chronic myelogenous leukemia. French Chronic Myeloid Leukemia Study Group. N Engl J Med 1997;337: (32) Hasford J, Baccarani M, Hehlmann R, Ansari H, Tura S, Zuffa E. Interferon-alpha and hydroxyurea in early chronic myeloid leukemia: a comparative analysis of the Italian and German chronic myeloid leukemia trial with interferonalpha. Blood 1996;87: (33) Kantarjian HM, Keating MJ, Smith TL, Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998 REPORTS 857

9 O Brien S, Beran M, Pierce S, et al. Prolonged survival in chronic myelogenous leukemia after cytogenetic response to interferon-alpha therapy. The Leukemia Service. Ann Intern Med 1995;122: (34) Hehlmann R, Heimpel H, Hasford J, and the German CML Study Group. Randomized comparison of interferon-alpha, hydroxyurea and busulfan in chronic myeloid leukemia: Response to Kantarjian and Talpaz and to Tura and Baccarani [letter]. Blood 1995;85: Notes The decisive financial support that enabled the Collaborative CML Prognostic Factors Project came from the German Science Foundation Wilhelm Sander-Stiftung. Additional support was contributed by Hoffmann-La Roche, Inc., and Essex-Pharma. The French study was supported by the Association Pour la Recherche sur le Cancer. We are grateful for the support of J. Guilhot (Hospital Jean Bernard, Poitiers, France), R. Prittwitz (Universitats Kliniken Essen, Essen, Germany), and S. Röder (Biometrisches Zentrum fur Therapiestuchen, Munich, Germany). Valuable comments were provided by Professor Dr. H. Heimpel (Medizinische Universitatsklinik Ulm, Ulm, Germany), P. Dirschedl (Institute for Medical Informatics, Biometry, and Epidemiology, University of Munich, Germany) and Professor Dr. H. Lyon (C. Everett Koop Institute at Dartmouth Medical School, NH). Manuscript received November 17, 1997; revised March 31, 1998; accepted April 6, REPORTS Journal of the National Cancer Institute, Vol. 90, No. 11, June 3, 1998