A20-03 AI-based Quantitative CT Analysis Links Baseline UIP Pattern to 10-year Progression of Interstitial Lung Abnormalities in the COPDGene Study

Abstract Rationale Long-term observational studies like COPDGene enable assessment of interstitial lung abnormality (ILA) progression. We previously showed that a deep learning-based classification of usual interstitial pneumonia (MIL-UIP) on baseline CT is associated with 5-year progression of ILA measured by quantitative CT (DTA). Here we extend this analysis to 10-year follow-up, evaluating associations between baseline MIL-UIP, longitudinal DTA progression, and survival. Methods We evaluated baseline, 5- and 10-year followup CT scans from COPDGene participants with ILA or indeterminate ILA at baseline using MIL-UIP and DTA. DTA scores were log-transformed and MIL-UIP was dichotomized using a cutoff of 0.5. We fit a linear mixed model for DTA which included terms for time (in years), MIL-UIP, and a time by MIL-UIP interaction. The model was adjusted for baseline age, BMI, height, sex, smoking status, and included random effects for study center, subjects, and slope for time. Additionally, to account for a CT protocol change aimed at reducing the radiation exposure at the 10 year visit, we included an estimate of CT image noise and an indicator variable for scanning protocol. We also compared the risk of mortality between MIL-UIP groups by fitting a multivariable Cox proportional hazards model with adjustments for baseline age, DTA, BMI, sex, and smoking status. Results In total, 3,829 participants (mean±SD age 60.5±9.4 years; 1,802 female) were included, with 87 (2.3%) participants who had MIL-UIP 0.5. For participants with MIL-UIP 0.5, the average relative annual increase in DTA was 10.4% (95% CI: 5.8%, 15.2%, p 0.001). This annual increase was significantly higher (p = 0.0064) compared to participants with MIL-UIP ≤ 0.5, who had an average relative annual increase in DTA of 4% (95% CI: 3.4%, 4.7%, p 0.001). In terms of survival, the hazard of death for participants with MIL-UIP 0.5 was 1.62 (95% CI: 1.17, 2.25, p = 0.0034) times greater than participants with MIL-UIP ≤ 0.5. Conclusions This study evaluates ILA progression over 10 years and shows that CT features detectable on baseline CT using deep learning-based methods are associated with greater increases in fibrotic abnormality and poorer outcomes. This abstract is funded by: NHLBI grants U01 HL089897 and U01 HL089856 and by NIH contract 75N92023D00011