The diagnosis and classification of asthma and COPD, which are the most common respiratory diseases, have been vital importance. Recently, machine learning has been great interest for processing health data due to this importance. The focus of previous studies has been to obtain the most predictive algorithm for diagnosing aforementioned diseases. However, variable importance has received little attention. The principal objective of this study is to investigate machine learning algorithms both predictivity and variable importance evaluation abilities. A comprehensive comparison has been carried out by considering different performance criteria including accuracy, precision, recall, Kappa statistics, F-measure, ROC curve and AUC value. The findings have been indicated that Random forest (with CART learner), C5.0 for asthma and SVM (with non-linear kernel) and GBM for COPD are the best algorithms. While the most important variable for asthma is FEV3, MEF50 for COPD. FEV1 and FVC have been common variables among the top five variables. The rank of variables has been supported by statistical significance tests. Additionally, ROC curves and AUC values have provided similar and supportive results as visually. Consequently, this study should be of value to practitioners and researchers studying on expert systems on health sciences and machine learning applications.

Username
Password