Feasibility of lung cancer prediction from low-dose CT scan and smoking factors using causal models

Abstract

Introduction: Low-dose CT (LDCT) is currently used in lung cancer screening of high-risk populations for early lung cancer diagnosis. However, 96% of individuals with detected nodules are false positives.

Methods: In order to develop an efficient early lung cancer predictor from clinical, demographic and LDCT features, we studied a total of 218 subjects with lung cancer or benign nodules. Probabilistic graphical models (PGMs) were used to integrate demographics, clinical data and LDCT features from 92 subjects (training cohort) from the Pittsburgh Lung Screening Study cohort.

Results: Learnt PGMs identified three variables directly (causally) linked to malignant nodules and the largest benign nodule and used them to build the Lung Cancer Causal Model (LCCM), which was validated in a separate cohort of 126 subjects. Nodule and vessel numbers and years since the subject quit smoking were sufficient to discriminate malignant from benign nodules. Comparison with existing predictors in the training and validation cohorts showed that (1) incorporating LDCT scan features greatly enhances predictive accuracy; and (2) LCCM improves cancer detection over existing methods, including the Brock parsimonious model (p<0.001). Notably, the number of surrounding vessels, a feature not previously used in predictive models, significantly improves predictive efficiency. Based on the validation cohort results, LCCM is able to identify 30% of the benign nodules without risk of misclassifying cancer nodules.

Discussion: LCCM shows promise as a lung cancer predictor as it is significantly improved over existing models. Validated in a larger, prospective study, it may help reduce unnecessary follow-up visits and procedures.

Keywords: cancer screening; low-dose CT; lung cancer risk.

Figure 1

The causal graph over all data of the training cohort. Nodes in the yellow box correspond to those directly associated with lung cancer status. A list of the variables used for this analysis is provided in online supplementary materials. Note that besides the edges represented by a direct arrow (A→B), all other edges do not exclude the possibility of a latent confounder. BMI, body mass index.

Figure 2

Comparison of MGM-FCI-MAX-derived with retrained lung cancer prediction models on the training cohort. (A) ROC curves were computed using nested 10-fold cross-validation. (B) Model discrimination measured by AUC. AUC, area under the ROC curve; BMI, body mass index; Ca, cancer; ROC, receiver operating characteristics.

Figure 3

(A) LCCM sensitivity/specificity plots of predictions across probability thresholds (validation cohort). (B) Distributions of predicted lung cancer score across models (validation cohort) for subjects with cancer (red) and benign nodules (blue). Brock parsimonious original refers to the model with the published coefficients. LCCM, Lung Cancer Causal Model.