Machine Learning Outperforms Traditional Quantitative Methods for CT-based COPD Diagnosis: An Analysis of the National Lung Cancer Screening Trial

Authors:

Abstract
Chronic obstructive pulmonary disease (COPD) remains underdiagnosed, even among patients undergoing lung cancer screening. Recent advances in deep learning, including task-specific convolutional neural networks (CNNs) and task-agnostic foundation models (e.g. Pillar-0), offer an opportunity to improve COPD risk prediction from low-dose CT and enable targeted spirometry within screening programs. We included 5,971 National Lung Screening Trial (NLST) participants with low-dose CT, spirometry, and clinical data, excluding scans with interstitial lung disease. COPD was defined as post-bronchodilator FEV₁/FVC <0.70. Three random forest risk models were developed by combining clinical variables with CT-derived imaging features: (1) a COPD-specific CNN probability, (2) Pillar-0 chest CT embeddings trained to predict COPD, or (3) Low attenuation area (LAA). Imaging models were trained on 10,957 COPDGene CTs and applied to NLST. Performance was assessed using area under the receiver operating characteristic curve (AUC). In participants ≥65 years, CNN- and Pillar-0–based risk models performed similarly in men (AUC 0.849 vs 0.827, p=0.21) and women (AUC 0.838 vs 0.823, p=0.48). In participants <65 years, CNN-based models outperformed Pillar-0 models in both men (AUC 0.859 vs 0.805, p<0.001) and women (AUC 0.861 vs 0.821, p<0.001). CNN and Pillar-0-based risk models outperformed all LAA-based risk models. These results show that machine learning improves COPD risk prediction vs LAA in lung cancer screening CT imaging. erj;68/suppl_70/PA2498/F1 F1 F1