Identification of clinically meaningful, overlapping obstructive respiratory disease subtypes via data-driven approaches in a primary care population

Abstract

Background: Obstructive respiratory conditions, including asthma, bronchiectasis, and chronic obstructive pulmonary disease (COPD), are increasingly recognised as heterogeneous syndromes with significant overlap. Multiple disease pathways contribute to phenotypes that do not always align with textbook definitions, limiting the effectiveness of a one-size-fits-all approach. This study aims to identify, validate, and characterise clinically meaningful airway disease subtypes using electronic healthcare records (EHR) and unsupervised machine learning clustering techniques.

Methods: We applied k-means clustering to 626,651 patients with a diagnosis of asthma, bronchiectasis, or COPD, using linked national structured EHRs in England. Twenty-one clinical features, including risk factors and comorbidities, were analysed, with dimensionality reduction via principal component and multiple correspondence analyses. Associations between cluster membership and exacerbations, as well as respiratory and cardiovascular mortality, were assessed. Over 3,696,962 person-years of follow-up, 102,522 deaths were recorded. Cluster stability was evaluated after five years, and genome-wide association studies (GWAS) were conducted to explore genetic associations with cluster membership.

Results: Seven clusters were identified, each encompassing patients across traditional diagnostic labels. Distinct clinical patterns emerged as follows: (1) High BMI female predominant, (2) Older male-predominant with diabetes and cardiovascular disease, (3) Eosinophilic atopic, (4) Older non-comorbid, (5) Non-comorbid low BMI, (6) Neutrophilic smoker, (7) Anxious/depressed female-predominant.The cluster with cardiovascular comorbidities showed the highest rates of hospital admissions for exacerbations. Neutrophilic cluster 6 is a potential novel subtype marked by persistent neutrophilia and poor outcomes. Cluster stability over five years ranged from 38% to 78%. GWAS revealed significant genetic loci in a cluster enriched for allergic disease and eosinophilia, suggesting shared genetic mechanisms.

Conclusions: This study provides a data-driven dissection of the heterogeneity underlying obstructive airway diseases in a large, real-world population. Unsupervised machine learning applied to national-scale EHR data revealed distinct and partially stable subtypes that transcend conventional diagnostic boundaries. These findings highlight the complexity and overlap of airway disease phenotypes and demonstrate the value of clustering approaches for uncovering clinically and biologically meaningful subgroups. This work lays the foundation for further exploration into mechanisms and prognosis within and across airway disease phenotypes.

Keywords: Asthma; Bronchiectasis; CALIBER; Chronic Obstructive Pulmonary Disease; Cluster Analysis; Electronic Health Records; Genome-wide association studies; Machine Learning; UKBiobank.

Fig. 1

Patient flow diagram. Individuals were eligible for inclusion if: a they were (or turned) 18 years of age or older during the study period, b they had been registered for at least one year in a primary care practice and c had a diagnosis of asthma, bronchiectasis or COPD, with the additional restriction that COPD diagnosis would have to be made age 35 or older

Fig. 2

Schematic illustration of the analysis pipeline

Fig. 3

Radal plot illustrating the average feature importance for some of the key cluster characteristics. Ranges and units for each feature are given in brackets

Fig. 4

Airways disease diagnosis broken down by cluster, sorted by asthma prevalence. k1: High BMI female-predominant, k2: Older male-predominant with Diabetes and CVD comorbidities, k3: Eosinophilic atopic, k4: Older non-comorbid, k5: Non-comorbid low BMI, k6: Neutrophilic smoker, k7: Anxious/depressed female-predominant

Fig. 5

Cumulative average number of exacerbations over a 5-year period: a in primary care and b in hospital

Fig. 6

Cumulative average number of Emergency Department attendance over a 5-year period: a Respiratory presenting complaint and b medical complaint (infectious disease, sepsis, cardiac conditions, cerebrovascular, respiratory, diabetes and endocrine)

Fig. 7

Alluvial diagram showing the crossover and stability of patients in each cluster over 5 years