Heterogeneity of prognostic profiles in non-small cell lung cancer: too many variables but a few relevant

Abstract

Objective: Many prognostic factors, exceeding 150, for non-small cell lung cancer (NSCLC) are mentioned in the literature. The different statistical weight of the some variables at issue, their heterogeneity and their clinical uselessness is reviewed.

Study design and setting: Survival analysis of a cohort of NSCLC operated (n = 1730, 1993-1997) was carried out utilizing different statistical approaches: Cox proportional hazard analysis (CPHA), logistic regression (LRA), and recursive partitioning (CART).

Results: CPHA identified 13 prognostic variables and 11 LRA. Of the 17 possible variables, 10 are coincident. CART provided five different diagnostic groups but only three differentiated survival levels. Parsimonious models were constructed including only T and N cancer staging variables. Areas under the ROC curve of 0.68 and 0.68 were found for CPHA and LGA parsimonious models respectively, and 0.72 and 0.71 for complete models.

Conclusion: Variables with a minimal impact on the respective models and thus with little or scarce predictive clinical repercussion were identified. Differences in the prognostic profile of survival can be caused by the different methodological approaches used. No relevant differences were found between the parsimonious and complete models. Although the amount of information managed is considerable, there continues to be a large predictive gap yet to be explained.