Predictors of outcome and treatment decisions of older patients with stage III non-small cell lung cancer

Abstract

19574 Background: In treating Stage III non-small cell lung cancer in the elderly, oncologists often empirically adjust treatment without clear guidelines. Identifying patient characteristics that impact prognosis will aid in creating better treatment algorithms in this patient population. Methods: A retrospective analysis was done on older patients (age > 70) treated for Stage III NSCLC (excluding “wet” IIIB) at the H. Lee Moffitt Cancer. Cox multivariate analysis identified variables impacting progression free survival, overall survival, treatment chosen, treatment interruptions, and hospitalizations. Correlation and regression tree analysis (CART) was performed to create potential decision making models. Results: 213 patients were evaluable. Multivariate analysis identified ECOG performance status (hazard ratio = 1.52, p = 0.005) and nodal status (HR = 1.36, p = 0.001) as negatively associated with PFS while BMI (HR = 0.96, p = 0.02) was positively associated. ECOG performance status (HR = 2.26, p < 0.0001), nodal status (HR = 1.18, p = 0.08), and CIRS3 (having a comorbidity classified as severity 3; HR = 1.33, p = 0.02) were negatively associated with OS. BMI (HR = 0.95, p = 0.002) and CIRSmean (mean CIRS-G severity score; HR 0.61, p = 0.01) were positively associated with OS. These variables were also identified as the most significant splitters, along with smoking status and pulmonary function, in CART analyses. Trees using PFS and OS as outcomes were created with receiver operating curves (ROC) ranging from 0.64–0.75. A CART analysis targeting treatment modality chosen had ROCs ranging from 0.49–0.78. Conclusions: Our multivariate analysis found ECOG performance status, nodal status, and severe medical comorbidity as negatively associated with survival. Unexpectedly, BMI (at initial treatment), independent of weight loss, was found to be positively associated. Also, the splitter variables identified by CART analysis were very similar to the general multivariate model, but the weights differed in the various analyses and subgroups of patients. The CART analyses suggest that we can create decision making models with a range of applicability comparable to those commonly considered useful for clinical guidelines, i.e. about 60% of straightforward application. No significant financial relationships to disclose.