MA03.09 Retrospective Predictive Performance of a Lung Cancer Screening Risk Prediction Model in a Clinical Lung Cancer Screening Program

Abstract

United States Preventive Services Task Force (USPSTF) and Centers for Medicare & Medicaid Services (CMS) recommendations for annual screening for lung cancer with low dose CT (LDCT) scans rely on age and smoking history to identify those at high risk for lung cancer. The Tammemagi et al. six year lung cancer risk prediction model, PLCOm2012, developed and validated in large lung cancer screening clinical trials, demonstrated good predictive performance in study participants. A 1.51% PLCOm2012 risk threshold has been reported to outperform CMS/USPSTF entry criteria. This is the first time the model predictive performance has been evaluated in clinical practice. Predictive performance of a reparameterized (no education predictor variable) six year lung cancer risk prediction model, PLCOm2012noEd, was retrospectively assessed in 2297 consecutive individuals that underwent clinical CT lung screening between January 1, 2012 and November 30, 2015. All patients met the National Comprehensive Cancer Network (NCCN) Lung Cancer Screening Guidelines Group 1 or Group 2 high-risk criteria. 79 cancers were detected in the 2297 screened individuals with a mean follow-up of 2.12 years (75.9% (60/79) – NCCN Group 1). The model six year mean risk for lung cancer was higher for participants with lung cancer, 4.71%, as compared to those without lung cancer, 3.54% (p=0.008). Area under the curve (AUC) of the receiver operator characteristics (ROC) was 0.635 (95% CI 0.577 – 0.693). At the 1.51% predicted risk recommended screening threshold overall sensitivity = 86.1%, specificity = 29.8%, and PPV = 4.2%. For NCCN Group 1 (similar to CMS/USPSTF entry criteria), sensitivity = 91.7%, specificity = 20.7% and PPV = 4.04%. For NCCN Group 2 (younger, lighter smoking history, no limit on time quit with one additional risk factor) mean predicted risk for participants with lung cancer was 2.35% as compared to 1.83% for those without lung cancer but the difference was not statistically significant; p=0.2374. As the incidence of lung cancer was the same in NCCN Group 2 and NCCN Group 1 (3.24% vs 3.51%; p=0.8566) the sensitivity of the model for NCCN Group 2 at the recommended 1.51% screening threshold was reduced to 68.4% with a specificity of 56.3%. Lung cancer risk prediction model, PLCOm2012noEd demonstrated reduced sensitivity in individuals meeting NCCN Group 2 high-risk criteria undergoing clinical CT lung screening and may not be appropriate to adequately assess risk of lung cancer in this population.