Associations of pulmonary parameters with accelerometer data

Abstract

In Europe it is estimated that the number of elderly people aged above 65 will have doubled by 2060. In several chronic pulmonary diseases patients can suffer recurrent exacerbation episodes that can lead to severe breathing or death. In this thesis we explore the association of physical activity to lung health parameters, focusing on cystic fibrosis and chronic obstructive pulmonary disease patients and a group of the general population. The main goals of the thesis were to assess the feasibility of classifying exacerbation episodes in cystic fibrosis and chronic obstructive pulmonary disease patients and to implement new parameters in the context of a cohort study.

We conducted four distinct studies involving in total over 250 subjects. We asked them to wear a set of accelerometers, including GT3X and RT3, recording physical activity for up to 14 days. The data was processed and several features extracted that were used as inputs in three different classification algorithms: logarithmic regression, neural networks and support vector machines.

We achieved an area under the curve of 67% with logarithmic regression, 83% with neural networks and 90% with support vector machines when classifying exacerbation episodes in chronic obstructive pulmonary disease. A neural network was achieved an accuracy of 85% distinguishing cystic fibrosis patients from healthy controls. We proposed, extracted and tested a large set of physical activity parameters for use in KORA-Age.

The work on classification of exacerbations in COPD patients is, to our knowledge, the first attempt based on features from accelerometer data. Overall SVM showed to be the most robust classifier with an area under the curve of 90%. Nevertheless the number of patients and episodes is too low to draw definitive conclusions. The next step to classify exacerbations in COPD is to design a study with a statistically significant number of exacerbation episodes.