A Prognostic Model for Advanced Stage Nonsmall Cell Lung Cancer. Pooled Analysis of North Central Cancer Treatment Group Trials

Transcription

1 781 A Prognostic Model for Advanced Stage Nonsmall Cell Lung Cancer Pooled Analysis of North Central Cancer Treatment Group Trials Sumithra J. Mandrekar, PhD 1 Steven E. Schild, MD 1 Shauna L. Hillman, MS 1 Katie L. Allen, BS 1 Randolph S. Marks, MD 1 James A. Mailliard, MD 2y, MD 2 James E. Krook, MD 3 Andrew W. Maksymiuk, MD 4 Kari Chansky, MS 5 Karen Kelly, MD 6 Alex A. Adjei, MD, PhD 1 James R. Jett, MD 1 1 Mayo Clinic College of Medicine, Mayo Clinic and Mayo Foundation, Rochester Minnesota. 2 Division of Oncology, Missouri Valley Cancer Center, Omaha, Nebraska. 3 Division of Oncology, Duluth Clinic, Duluth, Minnesota. 4 Division of Oncology, University of Manitoba, Winnipeg, Manitoba, Canada. 5 Southwest Oncology Group Statistical Center, Seattle, Washington. 6 Division of Oncology, University of Colorado, Denver, Colorado. Presented at the 2005 Annual Meeting of the American Society of Clinical Oncology, May 13 17, 2005, Orlando, Florida. Supported by the National Cancer Institute (grant CA-25224). y Deceased. Address for reprints: Sumithra J. Mandrekar, PhD, Division of Biostatistics, Mayo Clinic, Rochester MN, 55905; Fax: (507) ; mandrekar. sumithra@mayo.edu Received January 6, 2006; revision received February 22, 2006; accepted April 4, BACKGROUND. A pooled analysis was performed to examine the impact of pretreatment factors on overall survival (OS) and time to progression (TTP) in patients with advanced-stage nonsmall cell lung cancer (NSCLC) and to construct a prediction equation for OS using pretreatment factors. METHODS. A pooled data set of 1053 patients from 9 North Central Cancer Treatment Group trials was used. Age, gender, Eastern Cooperative Oncology Group performance status (PS), tumor stage (Stage IIIB vs. Stage IV), body mass index (BMI), creatinine level, hemoglobin (Hgb) level, white blood cell (WBC) count, and platelet count were evaluated for their prognostic significance in both univariate and multivariate analyses by using a Cox proportional-hazards model. RESULTS. Patients who had high WBC counts, low Hgb levels, PS >0, BMI <18.5 kg/m 2, and TNM Stage IV disease had significantly worse TTP and OS. Patients who had Stage IV disease with a high WBC count had a particularly poor prognosis. An equation to predict the OS of patients with Stage IV NSCLC based on pretreatment PS, BMI, Hgb level, and WBC count was constructed. CONCLUSIONS. In addition to the widely accepted prognostic factors of PS, BMI, and disease stage, both of the readily available laboratory parameters of Hgb level and WBC count were found to be significant prognostic factors for OS and TTP in patients with advanced-stage NSCLC. The authors prediction equation can be used to evaluate the benefit of a treatment in Phase II trials by comparing the observed survival of a cohort with its expected survival by using the patients own prognostic factors in place of comparisons with historic data that may have substantially different baseline patient characteristics. Cancer 2006;107: Ó 2006 American Cancer Society. KEYWORDS: multivariate modeling, nonsmall cell lung carcinoma, pooled analysis, prediction model. In 2006, lung cancer will cause an estimated 162,460 deaths in the U.S. 1 A substantial number of patients (39%) have distant tumor spread at diagnosis. 2 Although radiotherapy and chemotherapy are used in the treatment of many patients with locally advanced nonsmall cell lung carcinoma (NSCLC), multiagent chemotherapy is the standard treatment for fit patients who have advanced-stage NSCLC (TNM Stage IIIB with a positive pleural effusion and Stage IV). Unfortunately, the outcome of patients with advanced NSCLC generally is poor, and treatment appears to have a very modest effect on overall survival (OS). A large metaanalysis demonstrated a 2-month increase in median survival after platinum-based therapy and an absolute 10% improvement in the 1-year survival rate compared with best supportive care. 3 Although 2-drug regimens are bet- ª 2006 American Cancer Society DOI /cncr Published online 14 July 2006 in Wiley InterScience (

2 782 CANCER August 15, 2006 / Volume 107 / Number 4 ter than single agents, 3-drug regimens are no better than 2-drug combinations in terms of improving OS. 4 The thrust of current research has shifted to targeted systemic therapy that is hypothesized to target cancer cells selectively, thus reducing toxicity. 5 Trials of targeted agents, such as erlotinib and bevacizumab, have met with success in enhancing survival. 6,7 Although progress has been made, the survival of patients with advanced NSCLC remains poor, with a median survival of 6 to 12 months and 1-year survival rates between 20% and 50%. 3,7 We performed a pooled analysis of first-line, advanced-stage NSCLC North Central Cancer Treatment Group (NCCTG) trials to investigate and improve our understanding of the impact of several potential prognostic factors on OS and time to progression (TTP). This kind of pooled analysis allows us to assess the consistency of correlations (i.e., the magnitude and direction of the impact of prognostic factors on outcomes) across a large number of trials as opposed to the findings from 1 or 2 large trials. An equation to predict the OS of individual patients using their significant pretreatment factors was developed. In patients with advanced NSCLC in whom treatment benefits have been modest, it is important to evaluate accurately the efficacy of new therapies and to move forward only with treatments demonstrate a promising clinical benefit beyond what is predicted by a patient s own prognostic factors. Currently, the outcomes from singlearm, Phase II trials are compared with historic data, which may have substantially different baseline patient characteristics, resulting in differing expected survival from the cohort being evaluated. This mechanism of comparison has the potential to miss an effective treatment or to judge that an ineffective treatment is effective. Instead, a prediction equation based on patients own prognostic factors can be used to evaluate the benefit of a treatment in single-arm, Phase II trials by comparing the expected survival of a cohort with the observed survival, an approach that has been used successfully in other diseases. 8 This would improve our ability to assess the results of Phase II trials that play a critical role in designing large Phase III trials, which define the standards of care. MATERIALS AND METHODS Trial Characteristics Individual patient data were pooled from first-line NCCTG therapy trials for advanced-stage NSCLC (Stage IIIB with pleural effusion and Stage IV) that opened between 1985 and Seven patients who never received any study treatment were excluded from all analyses. Platinum-containing regimens were administered in 46% of patients and in 56% of NCCTG trials. See Table 1 for a detailed listing of the individual trial characteristics, including the Southwest Oncology Group (SWOG) randomized Phase III trial of paclitaxel plus carboplatin versus vinorelbine and cisplatin in patients with untreated advanced NSCLC (S9509) used in model validation Statistical Analysis OS was defined as the time from registration to death from any cause. TTP was defined as the time from registration to the first documented evidence of disease progression. Any patient who died without documentation of disease progression was censored at the time of death for the TTP analysis. The Cox proportional-hazards (PH) model was used for both univariate and multivariate analyses. The models were stratified by protocol, and analyses were performed on the data available based on the selected covariates for the respective endpoint. Gender, age, performance status (PS), and disease stage (prognostic factors that have a known impact on OS and TTP) were included in all multivariate models regardless of their significance. The pretreatment factors that were collected across most trials and were considered for inclusion in our multivariate model were body mass index (BMI), white blood cell (WBC) count, hemoglobin (Hgb) level, platelet (PLT) count, and creatinine level. Factors that were identified as significant predictors of OS and TTP in the univariate setting were categorized for ease of interpretation in the multivariate setting. In addition to the main affects, all 2-way interactions were evaluated. The effect of platinum-based therapy was explored by adding it as a variable to the final, unstratified, multivariate model. The power to detect various effects depends on the prevalence and number of levels of each prognostic factor. For a general guide, a sample size of 1053 patients provides at least 80% power to detect an effect reflected by a hazard ratio (HR) of 1.2 for a 2-level factor with a prevalence of 40% versus 60% (2-sided log-rank test; a ¼.05) using the actual accrual rates for the trials from 1985 to 2003 and assuming exponential survival with 2 years of minimum follow-up on each patient Age was categorized by using a cut-off age of 65 years (age 65 years vs. age <65 years), a commonly used cut-off age for defining the elderly, and was approximately the median age of patients who were in_cluded in this pooled analysis. A cut-off age of 70 years also was explored and resulted in similar findings. BMI was classified into 4 categories: underweight (BMI<18.5 kg/m 2 ), normal (BMI from 18.5 kg/m 2 to <25 kg/m 2 ), overweight (BMI25 kg/m 2 to BMI<30 kg/

3 Prognostic Model for NSCLC/Mandrekar et al. 783 TABLE 1 Individual Trial Characteristics * Protocol , y N S No. of patients Cancels Date opened 07/05/85 06/11/87 12/09/88 05/03/90 04/16/93 08/26/94 11/20/96 02/19/99 08/10/01 04/01/96 Date closed 08/21/90 08/17/88 03/15/91 10/23/92 11/11/93 06/04/96 05/15/98 01/26/01 05/27/03 01/01/98 Phase III III II III II II II II II III Performance status Stage IIIB/IV IV IV IIIB/IV IIIB/IV IIIB/IV IIIB/IV IIIB/IV IIIB/IV IIIB/IV Weight-loss criteria No exclusions No exclusions No exclusions Excluded >10% in 3 months Agents A) Mitomycin, vinblastine, and cisplatin; B) mitomycin A) Bolus etoposide plus cisplatin; B) infusion etoposide plus cisplatin A) Amonafide; B) trimetrexate A) cisplatin, etoposide, and hydrazine sulfate; B) cisplatin and etoposide Excluded >10% in <2 months Excluded >10% in <2 months A/B) Topotecan A) Cisplatin, topotecan, and filgrastim; B) topotecan, paclitaxel, and filgrastim BSA 2.3 Excluded >10% in <3 months A) EVAC/ filgrastim A/B/C) Docetaxel and gemcitabine Excluded >10% in <6 weeks A/B/C) Alimta and gemcitabine Factors collected Hgb Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes WBC Yes Yes Yes Yes Yes Yes Yes No { No { Yes PLT Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Creat Yes Yes Yes Yes No No No Yes Yes Yes BMI Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No exclusions I) Vinorelbine and cisplatin; II) paclitaxel, carboplatin BSA indicates body surface area (m 2 ); EVAC, edatrexate in combination with vinblastine, doxorubicin, and cisplatin; Hgb, hemoglobin; WBC, white blood cells; PLT, platelets; Creat, creatinine; BMI, body mass index. * Superscript numbers refer to the list of references. y Unpublished data { Absolute neutrophil count was collected instead of WBC. Creatinine was required to be within institutional normal limits for eligibility; actual values were not collected.

4 784 CANCER August 15, 2006 / Volume 107 / Number 4 m 2 ), and obese (BMI30 kg/m 2 ). 19 Gender-based cut-off points were determined based on the 2004 standards of the Mayo Clinic Reference Laboratory for WBC counts, Hgb levels, and PLT counts. The variable high WBC count was defined as a WBC count > /L for males and > /L for females. Similarly, the variable anemia (low Hgb) was defined an Hgb level <13.2 g/dl for males and <11.5 g/dl for females. High PLT was defined as a PLT count > /L for males and > /L for females. The OS and TTP models were tested for the appropriateness of the PH assumption. Any deviations from this assumption were explored further by the graphic inspection of Schoenfeld residuals. 20 The prognostic factors were tested for between-trial heterogeneity. Diagnostic plots and sensitivity analyses were conducted to explore the influence of individual patient data on the overall model outcomes. The final models were selected by using stepwise regression modeling techniques. Forward and backward regression procedures also were explored and provided similar results. All tests were 2-sided, with P values <.05 for main effects and P values <.01 for interaction terms denoting statistical significance. HRs and the associated 95% confidence intervals (95% CIs) were calculated. The final multivariate models for OS and TTP were assessed for internal validation by using bootstrap techniques (stratified samples with replacement from the original data set). The same model-development process outlined above was applied to each of the 1000 data sets to assess stability. 21 The percentage of times that a factor was selected as a significant factor in the multivariate model is reported. Factors that were selected in >60% of the analyses were considered stable. A prediction model for OS was developed that incorporated the significant pretreatment factors. The discriminative ability of the prediction model was evaluated by using the c-statistic. 22 This model was validated further independently by using the S9509 data set. 18 Data from both arms of the trial were combined, because there were no differences in OS between the 2 arms, and the arms were balanced for all of the prognostic factors considered. The calibration of our prediction model was assessed by comparing the predicted OS estimates with the corresponding observed Kaplan Meier estimates for subgroups with 20 patients. FIGURE 1. Patient cohorts included in the various analyses are represented. NCCTG indicates North Central Cancer Treatment Group; NSCLC, nonsmall cell lung carcinoma; WBC, white blood cell count; pts, patients; PLT, platelets; Hgb, hemoglobin; PS, performance status (Eastern Cooperative Oncology Group); ANC, absolute neutrophil count. RESULTS Data were frozen on February 11, 2005 and included 1053 patients in total (Fig. 1). Data were complete, with 96% of patients followed until death. The median OS ranged from 4.7 months to 10.8 months for all trials, and the median TTP ranged from 1.8 months to 4.9 months. The 1-year and 2-year OS and TTP estimates ranged from 6.1% to 44.5% and from 3.4% to 22.2%, respectively, across all trials. See Table 2 for the individual trial outcomes. Baseline Patient Characteristics Table 3 provides a detailed description of the patient characteristics for the 1053 patients, the cohort of 782 patients who were evaluable for the multivariate ana- TABLE 2 Trial Outcomes Protocol % Alive Median OS, months Median TTP, months 1-Year OS estimates (%) * N0026* Overall S OS indicates overall survival; TTP, time to progression. * Not included in the multivariate models. 2-Year OS estimates (%)

5 Prognostic Model for NSCLC/Mandrekar et al. 785 TABLE 3 Patient Characteristics Cohort 1 (n ¼ 1053)* Cohort 2 (n ¼ 782) y Cohort 3 (n ¼ 426){ Factor No. % No. % No. % Age, y Median (range) 64 (31 85) 64 (31 85) 62 (32 84) < ECOG performance status Gender Male Female Stage IIIB IV Missing Body mass index, kg/m 2 Median (range) 24.8 ( ) 24.6 ( ) 24.5 ( ) Obese Overweight Normal Underweight Missing White blood cells, 10 9 /L Median (range) 9.1 ( ) 9.1 ( ) 9.3 ( ) High Normal Missing Hemoglobin, g/dl Median (range) 13.3 ( ) 13.2 ( ) 13.1 ( ) Anemic Nonanemic Missing Platelets, 10 9 /L Median (range) 348 ( ) 355 ( ) 324 ( ) High Normal Low Missing Creatinine, mmol/l Median (range) 1.0 ( ) 1.0 ( ) 0.9 ( ) Missing k ECOG indicates Eastern Cooperative Oncology Group; ULN, upper limit of normal. * Cohort of patients used in the univariate analysis. y Cohort of patients used in the multivariate modeling. { Cohort of patients from the S9509 trial. Absolute neutrophil count was collected instead of white blood cells in 2 trials. k Creatinine was required to be within institutional normal limits for eligibility in 3 trials; actual values were not collected. lyses, and the cohort of 426 patients from the S9509 trial. The baseline characteristics of the patients who were included in the multivariate model and in the validation data set were similar to the characteristics of the patients from the entire cohort. Univariate Analysis BMI, WBC, Hgb, and PLT were found to be significant predictors of OS and TTP in the univariate setting and were assessed further in the multivariate setting. Creatinine was not a significant predictor for OS or TTP

6 786 CANCER August 15, 2006 / Volume 107 / Number 4 TABLE 4 Univariate Results for Overall Survival and Time to Progression (n ¼ 1053 Patients) * OS TTP Factor HR P 95% CI HR P 95% CI Age, y < Gender Female Male ECOG performance status < < < < Stage IIIB IV Body mass index Normal Underweight Overweight Obese White blood cells Normal High 1.48 < < Hemoglobin Nonanemic Anemic 1.59 < < Platelets Normal High 1.41 < Creatinine TTP indicates time to progression; OS, overall survival; HR, hazard ratio; 95% CI, 95% confidence intervals; ECOG, Eastern Cooperative Oncology Group. * These results were similar when repeated on the cohort of patients (n ¼ 782) used in the multivariate analysis. univariately and therefore was not explored further. See Table 4 for the univariate results. Multivariate Analysis---Main Effects All factors that were significant predictors in the univariate setting also were significant predictors in the multivariate setting except the PLT count, which was excluded from the final models. Patients with who had low Hgb levels (anemic), high WBC counts, low BMI (underweight), and a PS>0 had significantly worse TTP and OS. Age and gender were not found to be significant predictors for OS or TTP in the multivariate setting (Table 5). There were no significant differences in OS or TTP between patients who received platinum-based therapy versus others (platinum vs. no platinum) after adjusting for the significant factors in the multivariate model (OS: HR, 0.94; 95% CI, [P ¼.65]; TTP: HR, 0.97; 95% CI, [P ¼.86]). Figure 2A-C includes the predicted survival curves for BMI, Hgb level, and WBC count. These are visual representations of the magnitude of difference in survival expected for the different levels of a factor given a patient with advanced NSCLC who has Stage IV disease, a PS of 0, age 65 years or younger, male gender, a normal BMI, a nonanemic Hgb level, and a normal WBC count. The relative difference in the predicted survival curves does not change with a change in the fixed patient characteristics. An equation to predict OS for patients with Stage IV NSCLC that incorporated only the significant pretreatment factors (PS, BMI, Hgb level, and WBC count) from the final multivariate model was constructed. Specifically, the survival probability is estimated by: in which, x ¼ ([0.26] þ [0 if PS ¼ 0] or [0.48 if PS ¼ 1] or [0.96 if PS ¼ 2or3] þ [0.60 if underweight] or [0 if normal] or [0.11 if overweight] or [ 0.11 if obese] þ [0 if nonanemic] or [0.41 if anemic] þ [0 if normal WBC] or [0.35 if high WBC]). The estimates for S 0 (dependent on time) can be found in Figure 3. Model-derived estimates of 6-

7 Prognostic Model for NSCLC/Mandrekar et al. 787 TABLE 5 Multivariate Results for Overall Survival and Time to Progression TTP OS OS for S9509 Factor HR P 95% CI HR P 95% CI HR P 95% CI Age, y < Gender Female Male ECOG PS < < < , < < Stage IIIB IV BMI Normal Obese Overweight Underweight WBC Normal High < Hemoglobin Nonanemic Anemic 1.41 < < TTP indicates time to progression; OS, overall survival; HR, hazard ratio; 95% CI, 95% confidence intervals; ECOG, Eastern Cooperative Oncology Group; PS, performance status; BMI, body mass index; WBC, white blood cells. month and 1-year OS for the different patient subsets are given in Table 6. Multivariate Analysis---2-Way Interactions The only significant 2-way interaction was the disease stage-by-wbc interaction in the OS model (Fig. 4). This interaction suggests that the prognosis was particularly poor for patients with Stage IV NSCLC who had a high WBC count. The prognostic value of the WBC count in patients with Stage IIIB NSCLC was less clear. Although the predicted survival for patients with Stage IIIB disease who had a high WBC count appears to be relatively better, the number of patients in this subset was small (n ¼ 24 patients). Model Diagnostics and Validation All factors met the PH assumption. The variable Hgb level had an influential trial (trial ) (HR, 9.5) for OS; however, all trials showed an effect in the same direction with smaller magnitudes (HR range, ). Thus, including the Hgb level was justified, because the observed heterogeneity was only quantitative. For all other factors, the homogeneity assumption was satisfied. The c-statistic for the prediction model was 0.65, which reflects a moderate measure of discrimination. The addition of the PLT count to the final model changed the c-statistic by only 0.2%. The calibration of the model also was assessed internally by comparing the predicted 6-month OS rates with the corresponding observed survival for all subgroups with 30 patients. The majority of subgroups that were assessed (71%) were within 3 percentage points. The bootstrap results indicated that the modelselection process was very stable. The Hgb level, WBC count, BMI, and disease stage-by-wbc interaction were selected as significant predictors of OS in 98%, 98%, 88%, and 66% of analyses, respectively. The WBC count, Hgb level, and BMI were selected as significant predictors of the TTP in 95%, 94%, and 75% of analyses, respectively. The PLT count, which was not a significant predictor in the multivariate setting, was

8 788 CANCER August 15, 2006 / Volume 107 / Number 4 FIGURE 3. The predicted survival equation for patients with Stage IV nonsmall cell lung carcinoma. PS indicates performance status (Eastern Cooperative Oncology Group); overwt, overweight; underwt, underweight; WBC, white blood cell count. FIGURE 2. Predicted survival curves for (A) body mass index (BMI), (B) hemoglobin (Hgb), and (C) white blood cell count (WBC). selected in only 43% and 9% of analyses with the bootstrap models for OS and TTP, respectively. The multivariate model estimates for all factors were similar but were smaller in magnitude in the validation data set (except for age and BMI) compared with the estimates that were obtained from the pooled analysis data set (Table 5). Figure 5A,B show no observable differences in the observed and predicted survival estimates for the subgroups, whereas Figure 5C,D indicate that the observed survival of the subgroups was greater than the predicted estimates, suggesting that the model-based predictions were more precise for some subgroups than for others (see Table 6 for S9509 results in subgroups with 20 patients). DISCUSSION Brundage et al. 23 performed a systematic review of the literature by investigating patient and tumor factors that were predictive of survival for patients with NSCLC. Those authors concluded that individual studies typically were underpowered and remarkably heterogeneous in their conclusions. They recommended that larger studies with clinically relevant modeling were required to address the usefulness of prognostic factors in defining the management of patients with NSCLC. 23 Hoang et al. 24 developed a nomogram to predict the 1-year and 2-year survival estimates in patients with chemotherapy-naive NSCLC who received standard chemotherapy. Their analysis revealed the following negative prognostic factors: subcutaneous metastases, poorer PS, loss of appetite, liver metastases, >4 metastatic sites, and no prior lung surgery. Those authors did not explore hematologic parameters. To design future trials properly, it is critical that we understand the lessons learned from patients who already have received treatment for this disease. We have amassed a cohort of 1053 patients from 9 NCCTG trials with advanced stages of NSCLC who received first-line systemic therapy (mostly conventional chemotherapy). Patients with a poorer PS, Stage IV disease, a high WBC count, lower BMI, or anemia fared significantly worse in terms of TTP and

9 Prognostic Model for NSCLC/Mandrekar et al. 789 TABLE 6 Predicted 6-Month and 1-Year Overall Survival Estimates for Stage IV Patients * Overall survival estimates (%) Nonanemic patients Anemic patients Normal WBC High WBC Normal WBC High WBC Status 6 months 1 year 6 months 1 year 6 months 1 year 6 months 1 year Underweight BMI PS PS PS 2/ Normal BMI PS 0.70 (.87).45 (.58) PS 1.57 (.52).27 (.30).45 (.57).16 (.29).43 (.42).14 (.17).30 (.54).06 (.19) PS 2/ Overweight BMI PS 0.67 (.77).40 (.68) PS 1.53 (.68).23 (.32) (.33).04 (.19) PS 2/ Obese BMI PS PS PS 2/ WBC indicates white blood cells; BMI, body mass index; PS, performance status (Eastern Cooperative Oncology Group). * Numbers in parentheses reflect the observed survival rate in the validation data set for only those subgroups with 20 patients. OS. Of these, it is accepted generally that PS, lower BMI, and Stage IV disease are associated with poorer outcomes. 23 However, the prognostic importance of complete blood count findings has been reported less consistently None of the previous series observed that both the pretreatment Hgb level and the WBC count were of independent prognostic value. The results of our current review highlight the prognostic importance of both the pretreatment Hgb level and the pretreatment WBC count in determining OS and TTP. An interesting and previously unexplored finding of the current analysis is that the prognosis is particularly poor for patients with Stage IV NSCLC who have a high WBC count. This may reflect either a greater burden of tumor cells within the bone marrow, or a possible concomitant subclinical infection, or the effect of a yet undescribed chemokine or cytokine secreted by the tumor into the circulation. No significant differences in outcome were observed based on platinum therapy (platinum-containing regimen vs. others) after adjusting for the significant prognostic factors, indicating that the influence of the prognostic factors on OS was more profound than the benefit from platinum-based therapy. From a clinical perspective, a low Hgb level, a PS>0, and a BMI<18.5 kg/m 2 reflects the cachexia associated with cancer, especially metastatic cancer; and patients who develop cachexia usually do not survive very long. Therefore, when investigating new agents, it is critical to control (stratify) or adjust for these prognostic factors in the final analyses. Otherwise, these factors may confound the results of a trial and lead to misperceptions regarding the benefits of treatment. There were some limitations to the current analysis. Data for this analysis came primarily from Phase II trials that reported an overall median survival of 7 months, which is much shorter than the median survival observed in the S9509 trial (8.6 months) and in other recent Phase III trials. The overall sample size also was smaller compared to with an SWOG pooled analysis in a similar patient population. 25 Because data were pooled across 9 independent clinical trials, not all factors were collected in each trial. In addition, other factors of potential interest, such as the number and sites of metastases, histology, smoking status, and weight loss, were not collected routinely and, hence, were not evaluated. Some subgroups had relatively small numbers of patients available for evaluation and validation: notably, patients with Stage IIIB disease, patients with a low BMI, and patients who had Stage IIIB disease with a high WBC count. In addition, like

10 790 CANCER August 15, 2006 / Volume 107 / Number 4 FIGURE 4. Predicted survival curves for the white blood cell count (WBC) and stage interaction. in any clinical trial, only select proportions of patients participate in clinical trials; thus, the group of patients on whom the predicted survival estimates are based may not be representative of all patients with Stage IV NSCLC. Our external validation suggests that our model-based predictions certainly were more precise for some subgroups than for others. This may be because of differences in the effects of some prognostic factors, like age and disease stage, in the pooled sample versus the validation data set. However, PS, Hgb level, and WBC count were important prognostic factors in both data sets. Further refinement of the prognostic importance of pretreatment factors that were identified in this analysis will be necessary. Despite these limitations, the results from the current analysis are more representative than those from 1 or 2 large trials. The homogeneity assumption FIGURE 5. Observed and predicted survival estimates for the Southwest Oncology Group (SWOG) S9509 data set. (A) Subgroup: body mass index (BMI), normal; performance status (PS) (Eastern Cooperative Oncology Group), 1; white blood cell count (WBC), normal; and nonanemic. (B) Subgroup: BMI, normal; PS, 1; WBC, normal; and anemic. (C) Subgroup: BMI, overweight; PS, 1; WBC, normal; and nonanemic. (D) Subgroup: BMI, overweight; PS, 1; WBC, high; and anemic.

11 Prognostic Model for NSCLC/Mandrekar et al. 791 was satisfied for each prognostic factor, indicating that the magnitude and direction of the effect of the factor on OS was similar across trials. Thus, we demonstrated the consistency in the correlations observed over time and across trials, including a variety of treatments. The discriminative ability of the prediction model provided for a range of predictions; for example, a differential magnitude of benefit was observed in the 6-month and 1-year OS rates of 30% and 6%, respectively, for a patient with a normal BMI, a PS of 1, anemia, and a high WBC count compared with 6-month and 1-year OS rates of 43% and 14%, respectively, for a patient with a normal BMI, a PS of 1, anemia, and a normal WBC count. These are comparable to the rates published in the literature in which prediction models were developed for a population for the first time. 31 The usefulness of this work for clinical research and clinical practice is compelling in this population of patients with advanced NSCLC, in which advances in treatment have been slow with only modest improvements in survival. Data regarding WBC counts, Hgb levels, PS, and BMI are available readily for all patients, which makes our model a clinically feasible metric. In terms of clinical research, a prediction equation like that proposed here, based on the patients own prognostic factors, may be used to evaluate the benefit of a treatment in Phase II trials by comparing the observed survival with the expected survival of a cohort in place of historic comparisons, which may have substantially different baseline patient characteristics. This may allow a more accurate gauge of the value of new treatment programs prior to designing large-scale, randomized, Phase III trials, which are used to define the standards of care, as demonstrated in other diseases. 8 Specifically, at the end of a trial, the predicted survival probabilities for each patient in the trial, based on their own baseline values for each factor in the model, can be computed and compared with the observed survival to gauge the benefit from the current treatment above and beyond what is attributable to the patients prognostic (baseline) factors. Our model-based estimates also may be used upfront, when designing a trial, to choose the null hypothesis success rates. For a simple example, suppose a trial is designed for patients who have Stage IV NSCLC with eligibility criteria that require a normal baseline WBC count and Hgb level and, primarily, for patients with a PS of 1 who have a normal BMI. If the endpoint is the 6-month survival rate, then, based on the estimates given in Table 6, the predicted 6-month survival probability is 57% based on the given baseline factors. This trial can be designed to look for an improvement in this success rate, say, 15%. Thus, the model can help in setting up the null hypothesis rates based on the patient population expected to enroll in the trial rather than based on historic data. The existing Phase II trial designs for single-arm trials still will be applicable in this instance. Regarding the sample size for such a trial, it will depend on the null and alternative hypothesis rates and on other operating characteristics of the design, like the historic data approach. To our knowledge, this prediction equation is the first of its kind in a population with advanced-stage NSCLC and is reasonably accurate in for predicting the prognosis of patients with newly diagnosed, Stage IV NSCLC. The equation may be used to design trials and analyze results better. Further refinement and validation of this prediction equation will be needed to explore other factors of interest, such as metastatic status, as new prognostic information becomes available. REFERENCES 1. Jemal A, Siegel R, Ward E, et al. Cancer statistics, CA Cancer J Clin. 2006;56: American Lung Association. Available at: Accessed October 1, Non-Small Cell Lung Cancer Collaborative Group. Chemotherapy in non-small cell lung cancer: a meta-analysis using updated data on individual patients from 52 randomized clinical trials. Br Med J. 1995;311: Bunn PA Jr. Chemotherapy for advanced non-small-cell lung cancer: who, what, when, why. J Clin Oncol. 2002;20: 23s 33s. 5. Johnson DH. Targeted therapy in non-small cell lung cancer: myth or reality Lung Cancer. 2003;41:S3 S8. 6. Shepherd FA, Rodrigues Pereira J, Ciuleanu T, et al. Erlotinib in previously treated non-small-cell lung cancer. N Engl J Med. 2005;353: Sandler AB, Gray R, Brahmer J, et al. Randomized Phase II/ III trial of paclitaxel plus carboplatin with or without bevacizumab in patients with advanced non-squamous nonsmall cell lung cancer: an Eastern Cooperative Oncology Group trial: E4599. J Clin Oncol. 2005;23(16S):2s. Abstract LBA Solit DB, Morris M, Slovin S, et al. Clinical experience with intravenous estramustine phosphate, paclitaxel, and carboplatin in patients with castrate, metastatic prostate adenocarcinoma. Cancer. 2003;98: Veeder MH, Jett JR, Su JQ, et al. A Phase III trial of mitomycin C alone versus mitomycin C, vinblastine, and cisplatin for metastatic squamous cell lung carcinoma. Cancer. 1992;70: Castro M, Veeder MH, Mailliard JA, et al. A prospective study of pulmonary function in patients receiving mitomycin C. Chest. 1996;109: Goldberg RM, Jett JR, Therneau TM, et al. Bolus versus infusional etoposide and cisplatinum in treatment of nonsmall cell lung cancer: a study of the North Central Cancer Treatment Group. J Natl Cancer Inst. 1990;82:

12 792 CANCER August 15, 2006 / Volume 107 / Number Gesme DH Jr., Jett JR, Schreffler DD, et al. A randomized Phase II trial of amonafide or trimetrexate in patients with advanced non-small cell lung cancer: a trial of the North Central Cancer Treatment Group. Cancer. 1993;71: Loprinzi CL, Goldberg RM, Su JQ, et al. Placebo-controlled trial of hydrazine sulfate in patients with newly diagnosed non-small cell lung cancer. J Clin Oncol. 1994;12: Weitz JJ, Marschke RF, Sloan JA, et al. A randomized Phase II trial of 2 schedules of topotecan for the treatment of advanced stage non-small cell lung cancer. Lung Cancer. 2000;28: Colon-Otero G, Niedringhous RD, Hillman SH, et al. A Phase II trial of edatrexate, vinblastine, Adriamycin, cisplatin and filgrastim (EVAC/G-CSF) in patients with advanced non-small cell carcinoma of the lungs: a NCCTG trial. Am J Clin Oncol (CCT). 2001;24: Perez EA, Amin BR, Hillman SL, et al. Randomized Phase II study of docetaxel and gemcitabine for Stage IIIB/IV nonsmall cell lung cancer. Lung Cancer. In press. 17. Ma CX, Nair S, Thomas S, et al. Pemetrexed and gemcitabine as front-line therapy for advanced NSCLC: a randomized, Phase II NCCTG trial of three schedules. J Clin Oncol. 2005;23: Kelly K, Crowley J, Bunn P Jr., et al. Randomized Phase III trial of paclitaxel plus carboplatin versus vinorelbine plus cisplatin in the treatment of patients with advanced nonsmall-cell lung cancer: a Southwest Oncology Group trial. J Clin Oncol. 2001;19: Center for Disease Prevention. Available at: cdc.gov/nccdphp/dnpa/bmi/bmi-means.htm Accessed October 1, Grambsh PM, Therneau TM. Proportional hazards tests and diagnostics based on weighted residuals. Biometrika. 1994;81: Harrell FE Jr., Lee KL, Mark DB. Tutorial in biostatistics multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. 1996;15: Harrell FE, Lee KL, Califf RM, et al. Regression modelling strategies for improved prognostic prediction. Stat Med. 1984;3: Brundage MD, Davies D, Mackillop WJ. Prognostic factors in non-small cell. Lung cancer: a decade of progress. Chest. 2002;122: Hoang T, Xu R, Schiller JH, et al. Clinical model to predict survival in chemonaive patients with advanced non-smallcell-lung cancer treated with third-generation chemotherapy regimens based on Eastern Cooperative Oncology Group data. J Clin Oncol. 2005;23: Albain KS, Crowley JJ, LeBlanc M, et al. Survival determinants in extensive-stage non-small-cell lung cancer: the Southwest Oncology Group experience. J Clin Oncol. 1991; 9: Takigawa N, Segawa Y, Okahara M, et al. Prognostic factors for patients with advanced non-small cell lung cancer: univariate and multivariate analyses including recursive partitioning and amalgamation. Lung Cancer. 1996;15: Watine J. Further comments on A practical prognostic index for inoperable non-small-cell lung cancer : a clinical biologist s point of view. J Cancer Res Clin Oncol. 1998;124: Wigren T. Confirmation of a prognostic index for patients with inoperable non-small cell lung cancer. Radiother Oncol. 1997;44: Paesman M, Sculier JP, Libert P et al. Prognostic factors for survival in advanced non-small lung cancer: univariate and multivariate analyses including recursive partitioning and amalgamation algorithms in 1052 patients. J Clin Oncol. 1995;13: Ferrigno D, Buccheri G. Hematologic counts and clinical correlates in 1201 newly diagnosed lung cancer patients. Monaldi Arch Chest Dis. 2003;59: Gill S, Loprinzi CL, Sargent DJ, et al. Pooled analysis of fluorouracil-based adjuvant therapy for Stage II and III colon cancer: who benefits and by how much? J Clin Oncol. 2004;22: