In recent years, several researchers have attempted to identify COPD phenotypes using different cluster analysis. This study aimed to determine the most optimal cluster analysis (supervised vs unsupervised) to robustly identify clinical phenotypes. 2,968 COPD patients have been included from January 2014 until February 2020. General information (age, BMI, smoking, comorbidities), lung function, exacerbations and symptoms were collected. After 5 years of follow-up, vital status was recorded. A hierarchical classification on the principal components (HCPC) was performed, followed by two unsupervised classification algorithms: k-means and PAM (Partition Around Medoids). Robustness was defined according to three different indices of validation (Connectivity, Dunn and silhouette). The mean age was 70 years, 63.7% of males, current smokers: 38.7%, mean FEV1: 61.3% predicted, ≥2 exacerbations: 43.6%, mMRC dyspnea grade≥2: 56.3%, chronic cough: 58%. The 5-year mortality rate was 11.3%. Based on our hypothesis, four phenotypes were described, using the PAM method. The phenotype A (24.2%) consisted of elderly patients with severe airflow limitation, low symptoms, cardiovascular comorbidities, diabetes and a higher mortality. The phenotype B (23.9%) contained more female patients, young patients with moderate airflow limitation and a high rate of current smokers. The phenotype C (25.5%) contained patients with very severe airflow limitation, more symptoms and low BMI. The Phenotype D (26.2%) was composed of patients with mild airflow limitation and low dyspnoea. These results showed the superiority of PAM classification compared with two other algorithms (k-means and HCPC) in terms of the robustness.Read less <