P2.11-23 Performance of Lung Cancer Risk Prediction Models in I-ELCAP Smokers

Abstract

To determine the performance of lung cancer risk prediction models in predicting lung cancer in smokers enrolled in the International Early Lung Cancer Program (I-ELCAP). 62,071 asymptomatic ever-smokers enrolled into the international multi-institutions I-ELCAP for low-dose CT screening between 1993-2018. Demographics, smoking history, comorbidities, exposures and family history of lung cancer were collected at time of baseline CT scan. All participants received a baseline screening scan and subsequently annual repeat CT scans, and they were prospectively followed for the diagnosis of lung cancer. Diagnosis and treatment of lung cancer were verified and documented in the ELCAP Management System. To compare the predicted risk of lung cancer, we applied four lung cancer risk models: the Bach Model, the Liverpool Lung Project Incidence (LLPi) Risk model, the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial Model 2012 (PLCOM2012), and the Pittsburgh Predictor Model to the I-ELCAP cohort. Model calibration and discrimination were assessed using expected-to-observed (E/O) ratio and the area under the curve (AUC) statistics. E/O ratio >1 indicates that the model predicts more lung cancer cases than observed. PLCOM2012 model was the most predictive of lung cancer for ever-smokers in I-ELCAP with the AUC 0.61 being the highest, followed by Bach model (AUC 0.58), LLPi model (AUC 0.57) and Pittsburgh Predictor (AUC 0.52). E/O ratios suggested that the PLCOM2012 model, Bach model, LLPi model and Pittsburgh Predictor model tends to overestimate the number of lung cancers. The LLPi model overestimated as many as 4 times more lung cancer cases. Using data from I-ELCAP, the four existing lung cancer risk prediction models have AUCs ranged between 0.52-0.61, PLCOM2012 model was the top performer out of the four models.