Pulmonary nodule predictions: Model IDs likely malignancies

An analysis of two datasets from Canadian lung cancer trials has shown the rate of malignancy in lung nodules detected by low-dose CT screening to be less than 6 percent.

The results, published Sept. 5 in the New England Journal of Medicine, could help pave the way for predictive tools that use patient and nodule characteristics to estimate the probability that screening-detected lung nodules are malignant.

“An accurate and practical model that can predict the probability that a lung nodule is malignant and that can be used to guide clinical decision making will reduce costs and the risk of morbidity and mortality in screening programs,” wrote author Annette McWilliams, MB, of Vancouver General Hospital and the British Columbia Cancer Agency, and colleagues.

Low-dose CT screening in high-risk individuals has the potential to reduce lung cancer mortality, but at the risk of overdiagnosis and overtreatment. The landmark U.S. National Lung Screening Trial demonstrated a 20 percent drop in lung cancer mortality among those who were screened. However, the proportion of invasive diagnostic procedures in that study ranged as high as 4 percent, and there were 4.5 complications for every 10,000 people screened. One in four surgical procedures in the National Lung Screening Trial were performed on nodules that were later determined to be benign.

In an effort to improve the effectiveness of screening by better predicting which nodules are likely benign, McWilliams and colleagues looked at data from two low-dose CT screening cohorts: participants from the Pan-Canadian Early Detection of Lung Cancer Study (PanCan) and from trials at the British Columbia Cancer Agency (BCCA). The PanCan data set included 7,008 nodules and 102 malignancies across 1,871 participants, and the BCCA dataset featured 5,021 nodules and 41 malignancies in 1,090 participants. The PanCan and BCCA sets had rates of cancer of 5.5 percent and 3.7 percent, respectively.

Using the PanCan cohort as a development dataset for a predictive model, McWilliams and colleagues determined that predictors of cancer included:

  • Older age,
  • Female sex,
  • Family history of lung cancer,
  • Larger nodule size,
  • Upper lobe nodule location, and
  • Spiculation.

The authors then developed parsimonious and fuller multivariable logistic-regression models to estimate lung cancer probability, and used the BCCA data for validation. They found the models offered excellent predictive accuracy with receiver-operating-characteristic curves (AUC) of 0.94 in the external validation cohort.

“Even for lung nodules that were 10 mm or smaller, for which clinical management decisions are the most challenging, the AUC remained excellent (>0.90) in the validation cohort,” wrote the authors.

McWilliams and colleagues also found that the relationship between nodule size and cancer was nonlinear and that perifissural nodules represent a minimal risk of lung cancer that probably does not require CT follow-up.