Presentation (CTI)

OBJECTIVES: Associations between nonclinical characteristics and persistent exacerbation risk in patients with COPD are poorly defined, hindering targeted interventions for underserved groups. We used a machine learning-based clustering approach to characterise clinical, socioeconomic and demographic phenotypes among patients with uncontrolled COPD to identify predictors of poor outcomes.
METHODS: Using the US Guardian Research Network electronic medical record database (July 2021-June 2024), we identified 6,619 adults with uncontrolled COPD and complete data for analysis. Uncontrolled COPD was defined as ≥2 moderate exacerbations (outpatients requiring systemic corticosteroids/antibiotics) or ≥1 severe exacerbation (hospitalisation) within 12 months while receiving continuous double/triple inhaled maintenance therapy. LCA classified patients into mutually exclusive subgroups based on similarities in their multidimensional profiles (age, race, geography, area deprivation index [ADI], comorbidities, exacerbation history).
RESULTS: LCA identified four distinct phenotypes. Group 1 (16.3%), a socioeconomically vulnerable cohort characterised by 69% residing in deprived neighbourhoods and frequent moderate/severe exacerbations. Group 2 (12.0%), residing in socioeconomically disadvantaged neighbourhoods, with high representation of Black patients (42.9%) and relatively high frequency of moderate exacerbations but low hospitalisation rates. Group 3 (19.7%), older patients with lower ADI scores whose risk was associated with high comorbidity and hospitalisations. Group 4 (52.0%), predominantly receiving double therapy and with infrequent hospitalisation. Geographic analysis of the overall population identified clusters of patients with uncontrolled COPD in Missouri, South Carolina and Kentucky.
CONCLUSIONS: Patients with uncontrolled COPD have heterogeneous clinical, socioeconomic, demographic and geographic characteristics. For example, older populations with lower deprivation scores were characterised primarily by comorbidity burden and hospitalisation, while younger, more diverse patient populations exhibited frequent moderate exacerbations potentially associated with access-related barriers. Differentiating these groups is critical for advancing precision population health in COPD. These findings also demonstrate the value of integrating multidimensional clinical and nonclinical real-world data using machine learning approaches with the potential to inform targeted interventions.