Prediction of the Risk of Malignancy Among Detected Lung Nodules in the National Lung Screening Trial

Abstract

Objectives: This study aimed to investigate nodule features and patient-specific characteristics associated with improvement in predictive ability of lung cancer screening while maintaining the sensitivity of low-dose CT intact.

Methods: All authors were approved to use data from the National Lung Screening Trial, a previously conducted randomized clinical trial, through submission of a proposal to the Cancer Data Access System. The National Lung Screening Trial had a multilevel design with nodules nested within rounds and rounds nested within individuals; hence, to incorporate nodule-level features, multilevel logistic regression was used. Both nodule-level features and patient characteristics were included for model construction. Model construction was based on improvement in predictive ability of the model, and there were no restrictions to any significance level on variable inclusion.

Results: A total of 32,746 nodules for 9,728 patients were included in the analysis. With a sensitivity value equal to that of the National Lung Screening Trial (93.6%), positive predictive value was improved to 7.94%, which was more than twice that of the National Lung Screening Trial (3.6%). Area under receiver operating characteristic curve was 91.7% (95% confidence interval: 90.6-92.8).

Conclusions: Increment in positive predictive value of lung cancer screening with sensitivity same as National Lung Screening Trial is feasible, and inclusion of other nodule size dimensions plus longest diameter to the model significantly improves the predictive ability of models.

Keywords: Lung cancer screening; National Lung Screening Trial; lung CT scan; positive predictive value; prediction model; sensitivity.