Automatic Selection of the Most Characterizing Features for Detecting COPD in Speech

Speech can reveal important characteristics of a person such as accent, gender, age, and health. Identifying specific pathologies in a person’s speech can be extremely useful for diagnosis, especially if this can be done automatically. In the present research, we investigate which automatically computed speech features are characteristic for distinguishing Dutch COPD patients in exacerbated or stable condition. Read speech of a phonetically balanced story was recorded for COPD patients in exacerbation (n=11), stable condition (n=9), and for healthy controls (n = 29). Several acoustic features automatically computed with Praat and eGeMAPS were ranked by a Recursive Feature Elimination (RFE) method and were used as input for three binary classifications by both Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) classifiers: I exacerbation vs. healthy, II stable vs. healthy, and III exacerbation vs. stable. Besides the features for the full story, we also computed features on word and phoneme level. For all 9 combinations (3 binary comparisons 3 feature levels) we used RFE to select the top ten ranked features. The classification results showed: better performance for SVM (vs. LDA), better performance for story level (vs. word and phoneme level), and worse performance for comparison III exacerbation vs. stable (vs. I and II i.e. COPD vs. reference speech). A 100% correct classification could always be obtained at story level, but not for word and phoneme level; and only for a subset of features, not for all features. We discuss these results and consider their implications for future research and applications.