Establishing predictive models for malignant and inflammatory pulmonary nodules using clinical data and CT imaging features

Background: The detection of pulmonary nodules has become increasingly common; however, accurate qualitative diagnosis remains a clinical challenge. This study sought to distinguish between malignant and inflammatory solid lung nodules using clinical data and computed tomography (CT) imaging features. Methods: A total of 948 patients with pulmonary nodules who underwent surgery or percutaneous biopsy from four centers were included in the study. The patients were divided into the following four groups based on nodule diameter: Group 1: nodules <= 10 mm; Group 2: nodules >10 and <= 20 mm; Group 3: nodules >20 and <= 30 mm; and Group 4: all nodules. The independent risk factors were identified and merged by univariate and multivariate analyses in the four groups to establish four models. The overall performance of the four models was evaluated using the area under the curve (AUC) of the receiver operating characteristic curve. Differences between Models 1-3 and Model 4 were compared using the DeLong test. Results: Of the nodules, 638 were classified as malignant and 310 as inflammatory. The patients with malignant and inflammatory nodules had median ages of 64.3 +/- 9.8 and 56.0 +/- 11.9 years, respectively (P<0.001). To build the four models, 17 features were identified, of which 2 were clinical features and 15 were imaging features. Notably, the frequency of lobulation, age, multiple lesions, and satellite lesions was relatively high in the four models. The AUC, accuracy, sensitivity, and specificity of Models 1-4 were 0.861 (0.803-0.921), 73.5%, 81.0%, and 78.9%; 0.902 (0.873-0.931), 82.8%, 74.7%, and 88.0%; 0.943 (0.914-0.972), 90.5%, 87.3%, and 89.7%; and 0.921 (0.903-0.940), 84.7%, 83.1%, and 86.8%; respectively. However, there were no statistically significant differences between Models 1-3 and Model 4. Conclusions: Our novel subgrouping models were able to effectively distinguish between inflammatory and malignant lung nodules using a reduced feature set. Our models could facilitate the accurate diagnosis of patients with potentially malignant lesions.