A decision tree model to distinguish between benign and malignant pulmonary nodules on CT scans.

OBJECTIVE

Chest computed tomography (CT) is increasingly being used to screen for lung cancer. Machine learning models could facilitate the distinction between benign and malignant pulmonary nodules. This study aimed to develop and validate a simple clinical prediction model to distinguish between benign and malignant lung nodules.

PATIENTS AND METHODS

Patients who underwent a video thoracic-assisted lobectomy between January 2013 and December 2020 at a Chinese hospital were enrolled in the study. The clinical characteristics of the patients were extracted from their medical records. Univariate and multivariate analyses were used to identify the risk factors for malignancy. A decision tree model with 10-fold cross-validation was constructed to predict the malignancy of the nodules. The sensitivity, specificity, and area under the curve (AUC) of a receiver operatic characteristics curve were used to evaluate the model's prediction accuracy in relation to the pathological gold standard.

RESULTS

Out of the 1,199 patients with pulmonary nodules enrolled in the study, 890 were pathologically confirmed to have malignant lesions. The multivariate analysis identified satellite lesions as an independent predictor for benign pulmonary nodules. Conversely, the lobulated sign, burr sign, density, vascular convergence sign, and pleural indentation sign were identified as independent predictors for malignant pulmonary nodules. The decision tree analysis identified the density of the lesion, the burr sign, the vascular convergence sign, and the drinking history as predictors of malignancy. The area under the curve of the decision tree model was 0.746 (95% CI 0.705-0.778), while the sensitivity and specificity were 0.762 and 0.799, respectively.

CONCLUSIONS

The decision tree model accurately characterized the pulmonary nodule and could be used to guide clinical decision-making.