Prediction of true positive lung cancer in abnormal suspicious CT lung screens
Principal Investigator
Name
Martin Tammemagi
Degrees
DVM; MSc; PhD
Institution
Henry Ford Health System
Position Title
Professor of Epidemiology
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
201105-0015
Initial CDAS Request Approval
May 17, 2011
Title
Prediction of true positive lung cancer in abnormal suspicious CT lung screens
Summary
Computed tomography imaging of the chest is increasingly common. Abnormal imaging findings that are suspicious for lung cancer are common, often observed in over 25% of older individuals with a smoking history, and yet only a small proportion, <2%, turn out to be lung cancer. Clinicians and patients can make more informed clinical and personal decisions regarding a CT image that is abnormal suspicious for lung cancer if they have an accurate probability estimate of lung cancer. The current study plans to develop an accurate prediction model that will estimate the probability that an abnormal suspicious CT scan truly is lung cancer. A multivariable logistic regression model will be developed using NLST data that will provide probability of lung cancer given different CT abnormalities as well as sociodemographic, exposure, medical history and clinical data. This model will evaluate pulmonary nodules, but in addition will evaluate other abnormal suspicious lesions such as pulmonary mass, atelectasis, infiltration and hilar and mediastinal lymphadenopathy. Model predictive performance will be evaluated by assessing its discrimination using the receiver operator characteristic area under the curve and assessing calibration by evaluating the slope of the calibration line and the mean and 90th percentile absolute error between predicted and observed. Bootstrap correction of optimism or over-fit will be carried out to internally validate the model. This study will use existing data and does not require re-reading of CT images. It is expected that an accurate model will be produced which will have immediate utility in health practice.
Aims
(1) Our primary aim is to produce an accurate prediction model that will predict who truly has lung cancer amongst those individuals who have a CT lung cancer screen with an abnormal suspicious lesion for lung cancer.
(2) Our second aim is to internally validate the model’s predictive performance, including discrimination and calibration, using bootstrap methods.