Background: Chronic obstructive pulmonary disease (COPD) is a progressive disease affecting the airways, which constitutes a major cause of chronic morbidity and a significant economic and social burden throughout the world. Despite the fact that in COPD patients exacerbations are common acute events causing significant and often fatal worsening of symptoms, an accurate prognostication continues to be difficult. Objectives: To build computational models capable of distinguishing between normal life days from exacerbation days in COPD patients, based on physical activity measured by accelerometers. Methods: We recruited 58 patients suffering from COPD and measured their physical activity with accelerometers for 10 days or more, from August 2009 to March 2010. During this period we recorded six exacerbation episodes in the patients, accounting for 37 days. We were able to analyse data for 52 patients (369 patient days), and extracted three distinct sets of features from the data, one set of basic features such as average, one set based on the frequency domain and the last exploring the cross-information among sensors pairs. These were used by three machine-learning techniques (logarith mic regression, neural networks, support vector machines) to distinguish days with exacerbation events from normal days. Results: The support vector machine clas - sifier achieved an AUC of 90% ± 9, when supplied with a set of features resulting from sequential feature selection method. Neural networks achieved an AUC of 83% ± 16 and the logarithmic regression an AUC of 67% ± 15. Conclusions: None of the individual feature sets provided robust for reasonable classi - fication of PA recording days. Our results indicate that this approach has the potential to extract useful information for, but are not robust enough for medical application of the system.