Clinical Trajectory Analysis With Longitudinal Validation in COPD: A COPDGene Study

RATIONALE: Chronic obstructive pulmonary disease (COPD) is heterogeneous in its clinical phenotypes (e.g. chronic bronchitis, emphysema) and trajectories of disease progression. Analysis of large high-dimensional datasets presents a key opportunity to address the gap in our understanding of COPD phenotypes and progression. Clinical trajectory analysis (ClinTrajAn), based on the concept of the branching principal tree, simultaneously phenotypes and determines patient trajectories within cross-sectional clinical data. Our aim was to apply ClinTrajAn to map prominent subtypes and trajectories in a large population of participants, covering the whole range of COPD severity and at-risk profiles, and validate proposed trajectories using longitudinal data. METHODS: Cross-sectional data for 8972 participants from Phase 1 of the COPDGene longitudinal study were utilized for model training, with 4585/8972 (51%) of participants having Phase 2 data (∼5 years later). Participants included current and former smokers with COPD (GOLD 1-4), normal spirometry (GOLD 0), and preserved ratio impaired spirometry (PRISm). 30 features were selected for training, covering demographics, exposure, pulmonary function, and CT imaging. The Phase 1 data matrix (8972x30) contained 2302 missing values (< 1%), which were imputed via single value decomposition (SVD). Principal component analysis (PCA) was applied to this completed matrix to reduce dimensionality to the first six principal components. A bifurcating principal tree fitting this reduced data was computed by averaging over 100 iteratively grown trees fitting random 95% samples. Longitudinal displacement was determined via projection of SVD imputed Phase 2 data using Phase 1 PCA results. RESULTS: The averaged tree contained six terminal segments and two notable bridging segments (Figure 1 A). Terminal segments divided emphysema dominant COPD by sex, identified mild-to severe COPD participants with bronchodilator reversibility (BDR), chronic bronchitis dominance, healthy aged participants, and PRISm dominance. Bridging segments divided healthy aged and PRISm participants from COPD, and mild COPD or chronic bronchitis from severe COPD or participants with COPD and BDR. Trajectories were defined as paths starting from a root among GOLD 0 participants. Longitudinal analysis showed most participants (69%) stayed on the same segment after 5 years, with segment displacements on average moving away from the root, and a notable increase in displacement for cases with accelerated decline leading to a COPD subtype or PRISm terminal (Figure 1 B). CONCLUSIONS: We have applied ClinTrajAn in a large longitudinal study population to model phenotypes and trajectories in COPD, and validated prediction of progression pathways through observation of projected displacements over 5 years.