Abstract Rationale Chronic lung diseases, including chronic obstructive pulmonary disease (COPD) and interstitial lung disease (ILD) are frequently diagnosed late in the disease course. Structural abnormalities on chest computed tomography (CT) can be measured through quantitative imaging analysis (QIA) and are correlated with physiologic lung function in research cohorts; the utility of QIA metrics in real-world cohorts has been incompletely explored. We hypothesize the integration of QIA measures through multivariate machine-learning algorithms will have utility in predicting emphysema, obstruction, and restriction in a real-world lung cancer screening (LCS) cohort. Methods A subset of individuals enrolled in VA Boston’s LCS program who completed a low-dose CT (LDCT) from 4/2019-4/2024 were included; QIA-features were measured using Chest Imaging Platform software and were integrated with pulmonary function tests extracted from the medical record (training cohort) or obtained during a research study visit (replication cohort). Using the training cohort, we applied Distributional Random Forest (DRF) to jointly model multiple outcomes. Predictors were QIA features from the upper, middle, and lower lung thirds, and outcomes were FEV1/FVC and CT-derived total lung volume. From the model output, we derived a composite QIA score, which was tested in the replication cohort for associations with quantitative emphysema (%LAA −950 HU, Perc15), TLC (% predicted), spirometry-defined obstruction (FEV1/FVC 0.7), and restriction (TLC 80% predicted), adjusting for demographic, smoking, and scanner variables. Benjamini-Hochberg p-values 0.05 were considered significant. Measurements and Main Results Training cohort individuals (n = 535) were older, had higher pack-years of smoking, and more cardiometabolic comorbidities (e.g., diabetes, coronary disease), lower FEV1, and more quantitative emphysema relative to the replication cohort (n = 172). Applying DRF to the training sample, the most influential QIA predictors were from the left lower and right middle lung regions. In the replication cohort, higher composite QIA-scores were associated with lower quantitative emphysema (Perc15 β = 27.3 95% CI 25.3-29.3), lower TLC% predicted (β = -10.4 -11.7 to -9.13), less obstruction (OR 0.52 0.42-0.65), more restriction (OR 5.0 3.63-7.0), and reticular markings (OR 1.7 1.3-2.21). Visualization of composite QIA-scores plotted against quantitative emphysema (% LAA) and TLC% distinguishes emphysema-predominant, restricted-predominant, and combined obstructive-and-restricted subgroups (Figure 1). Conclusions Distributional random forest techniques which integrate QIA-derived metrics from real-world, clinically-available LCS LDCT data have the potential to identify distinct chronic lung disease phenotypes. Figure 1. Composite QIA-score plotted against % LAA -950 HU and TLC % predicted identified distinct subgroups with emphysema-predominant, restrictive-predominant and combined obstructive-and-restrictive clusters. This abstract is funded by: Department of Veteran Affairs
Moll et al. (Fri,) studied this question.