Validation of survival prognostic models for non-small-cell lung cancer in stage- and age-specific groups

Abstract

AIMS: Performance status (PS) is commonly used for patient stratification in lung cancer clinical trials. Many prognostic models have been proposed to predict survival for non-small-cell lung cancer (NSCLC). It is of interest to evaluate whether the existing prognostic models perform better than PS alone in stage- and age-specific trials for patient stratification. For stage- and age-specific groups, we aim to assess the performance of several popular prognostic models for overall survival with NSCLC trial data from Cancer Leukemia Group B (CALGB). PATIENTS AND METHODS: Data from Alliance/CALGB were used to validate the following prognostic models. Using 2979 NSCLC patients, B model (Blanchon et al 2006) was developed on a Cox regression with age, histology, PS, sex and stage (I-IV) as prognostic variables. Using 9137 surgically resected NSCLC patients, C model (Chansky et al 2009) was developed via Cox regression and recursive partitioning and amalgamation analyses with prognostic variables age, sex and stage (I-IIIA). Using 392 early stage NSCLC patients, G model (Gail et al 1984) was based on Weibull survival model with histology, PS and TN stage as prognostic variables. Using 782 advanced NSCLC patients, M model (Mandrekar at al 2006) was developed via Cox regression with prognostic variables age, BMI, hemoglobin level, PS, sex, stage (IIIB-IV), white blood cell count. In validation analysis, 1921 stage I (IA-IB) and 1108 stage IV NSCLC patients who participated in completed CALGB treatment trials and lung cancer tissue bank were included. For stage I, B, C and G models were evaluated along with the PS only model. For stage IV, B and M models were evaluated. The concordance of predicted survival times and risk scores was estimated by c-index (Harrell et al 1996). Separation of survival curves between risk groups was examined using Kaplan-Meier method and log rank test. The assessments were conducted for young (<70) and old (>=70) age groups as well as all patients combined. RESULTS: For stage 1 and young, B and PS have better survival separation than C and G with c-index for B, PS, C and G equal to 0.587, 0.567, 0.546 and 0.469. Slightly less concordant but similar findings hold for stage I and old patients with B, PS, C and G equal to 0.607, 0.586, 0.543 and 0.485. For stage IV and young, B and PS perform better than M with c-index for B, PS and M equal to 0.616, 0.594 and 0.556. Again, slightly less concordant but similar findings hold for stage IV and old patients with B, PS and M equal to 0.571, 0.552 and 0.492. CONCLUSION: Overall, B model performs better than other models for both stage I and stage IV and each age group. However, none of these prognostic models yield sufficiently better survival separation and rank concordance than the PS only model. Further research is needed to develop prognostic models significantly better than performance status alone for patient stratification in stage- and age-specific clinical trials. ©2013 American Association for Cancer Research