Analysis of Clinical and Radiologic factors associated with follow up Lung RADS score and lung cancer diagnosis in a low-intermediate risk lung cancer screening population

Abstract

Background Lung nodules identified by Low-Dose Computed Tomography (LDCT) are classified radiologically in the Lung Reporting and Data System (LungRADS). Herein, we elucidate clinical and radiomic factors associated with follow-up LungRADS 4 categorization and lung cancer diagnosis in patients with initial LungRADS 2-3 nodules. Methods We reviewed 2847 patients’ charts with LDCT screening between 2015 and 2019, identifying 2174 (76.7%) with LungRADS 2-3 baseline nodules, of which 908 had follow-up screening available (41.8%). We abstracted gender, race, smoking history, body mass index (BMI), history of chronic obstructive pulmonary disease (COPD), emphysema, pneumonia, cancer, and family history of lung cancer. Association with subsequent LungRADS 4 nodules and lung cancer was assessed via Cox proportional hazards regression. Among a subset of 132 patients, we performed radiomic analysis by annotating the largest nodule of each baseline scan and training machine learning models using 86 textural features. Clinical, radiomic and combined models were developed via Lasso regression and random forest analyses using five-fold cross-validation to predict a subsequent LungRADS 4 result or lung cancer diagnosis. Results Among patients with baseline LungRADS 2-3 nodules, 23 (1.1%) developed lung cancer. Age(p=0.04), pack-years (p<0.005), family history of lung cancer (p<0.005), COPD (p=0.02) and emphysema (p<0.005) were associated with lung cancer. 39 patients with follow-up screening (4.3%) progressed to LungRADS 4. Emphysema (p<0.005) was associated with this finding. Machine learning models trained with clinical variables yielded an average area-under-receiver operating characteristic curve (AUC) of 0.543 (random forest) and 0.799 (Lasso regression) for subsequent findings of LungRADS 4 and lung cancer. Radiomic analysis of the largest baseline nodule found an AUC of 0.574 (random forest) and 0.728 (Lasso regression) for findings of LungRADS 4 and lung cancer. Fusion of clinical and radiomic features yielded AUCs of 0.591 (random forest) and 0.718 (Lasso regression) for findings of LungRADS 4 and lung cancer. Conclusion Clinical and radiomic factors showed association with lung cancer diagnosis and LungRADS 4 categorization on follow-up screening in patients with baseline LungRADS 2-3 nodules. Assessment of these features may aid to predict which patients with likely benign screening findings may progress to a higher LungRADS category or develop lung cancer.