This study was undertaken to assess the validity of cluster analysis for stratifying patients with severe COLD into homogenous subgroups in view of further prospective studies. To this aim, physiological measurements and questionnaire data were obtained from 532 outpatients with severe COLD (e.g. a 1 sec forced expiratory volume (FEV1) below 1.5-1/sec). The model variables selected for the partition in cluster were FEV1, PaO2, response to bronchodilators and heart rate. Two subgroups of patients were identified by the analysis: cluster I with significantly greater physiological impairment than cluster II. The comparison of the prevalences of the variables outside the model between the 2 clusters showed, in fact, that cluster I had a significantly higher prevalence of subjects with heavy smoking (p less than 0.01), prolonged occupational exposure (p less than 0.05), low body weight (p less than 0.05), recent hospitalizations for respiratory troubles (p less than 0.02) and emphysema (p less than 0.01). In conclusion, cluster analysis based on few physiological variables was able to identify, among patients with severe COLD, those with poorer general conditions and higher exposure to specific risk factors, for whom a worse prognosis of life can be expected. The advantages of cluster analysis in comparison to other techniques of classification in this kind of patient is discussed.