Using predictive models to determine the presence of non-small cell lung cancer metastasis to N2 and N3 regions.

Abstract

e20560 Background: Accurate assessment of non-small cell lung cancer (NSCLC) mediastinal involvement is key to developing treatment plans and determining prognosis. To date, there is no reliable imaging-based means to determine the presence or absence of mediastinal involvement. Current computed tomogram (CT) and fluorodeoxyglucose-positron emission tomography/ computed tomogram (PET-CT) technologies provide numerous derived automated variables have not been sufficiently evaluated to determine the presence of metastasis to the mediastinum. We have developed predictive models to determine the presence or lack of metastatic NSCLC in N2 and N3 regions. Methods: Consecutive patients from 2012-2017 with biopsy-proven NSCLC who had CT and PET-CT, as well as biopsy of the mediastinum had their images reread by a team of blinded specialty radiologists and nuclear medicine specialists. Patients with no mediastinal malignancy on biopsy were followed for 6 months from the initial evaluation to confirm lack of mediastinal malignancy.278 regions (N2 and N3) from 139 patients were included. Logistic regression models were used to build a baseline model, as well as models with additional nodal station maximum standard uptake valuve (SUVm) measurements (SUVm, SUVm-SUVmeanbloodpool and SUV lymph node/tumor (LN/T)) for N2 and N3 regions, respectively. When nodal station SUVm was not measured, SUVmeanbloodpool was used. The SUVm within each region was used. Stepwise selection was used to select variables in the baseline model. Cross-validated ROC curve and area under the curve (AUC) were reported. All analyses were done in SAS 9.4 (SAS Institute, Inc., Cary, NC). Results: 40/139 N2 regions had malignancy, 4/139 N3 regions had malignancy. Baseline models for N2 regions selected lung laterality (OR right vs left: 4.84 (1.79, 13.05)) and nodal station short-axis diameter > 1 cm (OR yes vs no: 5.49 (1.71, 17.54)) while no variables were selected for the baseline model for N3 regions due to lack of statistical power. We used the same variables for the N3 baseline model. Conclusions: We have identified models that use a more advanced analysis of predicting the presence or absence of metastatic NSCLC in both N2 and N3 regions with respect to the primary lesion. All models perform better with SUVm related measurements. From this information, we are developing a clinical application to provide practitioners a better means of assessing the presence of mediastinal involvement of NSCLC. [Table: see text]