Contribution to Cluster Analysis in Chronic Obstructive Airway Diseases

Chronic obstructive pulmonary disease (COPD) and asthma are complex, multidimensional, and heterogeneous diseases, that represent an important burden for the public health expenditure in western world. In literature, patients are divided into several different groups according to the combination of clinical, biological, and physiological characteristics, and these groups are called phenotypes. Understanding the phenotype of each patient is the first step toward effective personalized management and treatment. The common statistical approach to determine the phenotypes is cluster analysis. Cluster analysis is a well-known unsupervised learning methodology that considers multiple variables in order to create coherent subsets among a large group of patients.

In this thesis, one of the most competitive and complex statistical analysis frameworks for applying cluster analysis in incomplete large datasets was introduced. In this framework, in addition to handling the missing values by multiple imputation, the dimensions of variables were reduced, and after performing the clustering method, the final result of clustering was achieved using a novel and efficient mixture multivariate multinomial model (4M) method. The efficiency of the proposed framework was evaluated and compared using several scenarios on simulated datasets with different competitive methods for each step.

The new framework was applied to three novel specific populations of COPD and asthma. The first study was conducted on 178 stable COPD patients with the ratio of forced expiratory volume in one second (FEV1) to forced vital capacity (FVC) post bronchodilation less than 70%, age above 40 years, and smoking history of at least 20 pack years and no clinical history of asthma before the age of 40 years. As a result, three different clusters were found, which shared similar smoking history. Including markers of systemic and airway inflammation and atopy and applying a comprehensive cluster analysis, we provide here evidence for 3 clusters markedly shaped by sex, airway obstruction, and neutrophilic inflammation but not by symptoms and T2 biomarkers.

In the next study, 426 eosinophilic patients which were defined by a sputum eosinophil count >=3% were considered. On the whole cohort, cluster analysis revealed two groups identified as cluster 1 (n=276) and cluster 2 (n=150) with cluster 1 being highly atopic with achievable control of the disease with ICS in most of the cases whereas cluster 2 featured a more aggressive disease, largely non-atopic with mixed granulocytic inflammation often resisting to ICS or oral Corticosteroids (OCS).

Finally, the framework was applied to a large group of asthmatics (n=588) who were non-eosinophilic (sputum eosinophils <3%). The analysis of the whole cohort revealed two groups identified as cluster 1 (n=417) and cluster 2 (n=171) with cluster 1 displaying a low treatment burden and proportion of atopy, a neutrophilic airway inflammation, a frequent smoking history with preserved lung function but poor asthma control and quality of life while the cluster 2 essentially featured atopic patients with paucigranulocytic and partly controlled asthma.

In conclusion, our proposed framework has an effective performance compared to competing methods based on the designed scenarios on these simulated datasets. By including airway inflammatory parameters among the variables, we have provided original data on cohorts of COPD and eosinophilic and non-eosinophilic asthmatics, which indicate substantial heterogeneity between clusters and, in asthma, in particular, great differences inside each airway inflammatory phenotype. Our findings should be confirmed in multicentric studies and their clinical value assessed on longitudinal studies looking at mortality and hospitalization in COPD and exacerbation rate and lung function decline in asthmatics.