Lung cancer risk prediction: Prostate, Lung, Colorectal And Ovarian Cancer Screening Trial models and validation

Abstract

Introduction: Identification of individuals at high risk for lung cancer should be of value to individuals, patients, clinicians, and researchers. Existing prediction models have only modest capabilities to classify persons at risk accurately.

Methods: Prospective data from 70 962 control subjects in the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO) were used in models for the general population (model 1) and for a subcohort of ever-smokers (N = 38 254) (model 2). Both models included age, socioeconomic status (education), body mass index, family history of lung cancer, chronic obstructive pulmonary disease, recent chest x-ray, smoking status (never, former, or current), pack-years smoked, and smoking duration. Model 2 also included smoking quit-time (time in years since ever-smokers permanently quit smoking). External validation was performed with 44 223 PLCO intervention arm participants who completed a supplemental questionnaire and were subsequently followed. Known available risk factors were included in logistic regression models. Bootstrap optimism-corrected estimates of predictive performance were calculated (internal validation). Nonlinear relationships for age, pack-years smoked, smoking duration, and quit-time were modeled using restricted cubic splines. All reported P values are two-sided.

Results: During follow-up (median 9.2 years) of the control arm subjects, 1040 lung cancers occurred. During follow-up of the external validation sample (median 3.0 years), 213 lung cancers occurred. For models 1 and 2, bootstrap optimism-corrected receiver operator characteristic area under the curves were 0.857 and 0.805, and calibration slopes (model-predicted probabilities vs observed probabilities) were 0.987 and 0.979, respectively. In the external validation sample, models 1 and 2 had area under the curves of 0.841 and 0.784, respectively. These models had high discrimination in women, men, whites, and nonwhites.

Conclusion: The PLCO lung cancer risk models demonstrate high discrimination and calibration.

Figure 1

Receiver operator characteristic (ROC) plots for models 1 and 2. A) Model 1, developed in and applied to all PLCO control subjects; B) model 1 applied to external validation dataset (PLCO intervention arm); C) model 2 developed in and applied to smokers in the PLCO control arm; D) model 2 applied to smokers in external validation set (PLCO intervention arm). PLCO = Prostate, Lung, Colorectal, Ovarian cancer screening trial.

Figure 2

Estimated probabilities of developing lung cancer in smokers for four predictors in unadjusted logistic regression models. A) Age. B) Pack-years smoked. C) Smoking quit-time. D) Smoking duration. The nonlinear relationships between these predictor variables and lung cancer risk were estimated using restricted cubic splines. Splines for age, pack-years smoked, quit-time and smoking duration were prepared with knot placement based on the percentile distributions of these variables in smokers only. Knots for age were at 55, 60, 64, and 72 years. Knots for pack-years were at 3.25, 23.25 and 63 pack-years. Knots for quit-time were at 0, 15, and 35 years. Knots for duration were at 8, 28, and 45 years. The y-axis in each of the four figures is the same scale to allow comparison of the relative impacts of each predictor on lung cancer risk.