PATTERN RECOGNITION APPLIED TO COUGH CATEGORIZATION
TEXAS UNIV AT AUSTIN LABS FOR ELECTRONICS AND RELATED SCIENCE RESEARCH
Pagination or Media Count:
The particular problem with which the research was concerned was the development of a technique to discriminate between coughs and other audible phenomena which originate in a hospital environment. Pattern recognition provided such a technique. Experimental data was available in the form of audio tape recordings. Implementation of a Bayes categorization decision requires knowledge of the underlying conditional joint probability density functions of the measures which typify the patterns to be recognized. An adaptive pattern classifier model was presented which circumvented the difficulty of estimating these functions. The model is generally applicable to the two-class case in which the patterns to be classified consist of sequential segments of data known to have originated from the same class. The model took the form of a layered machine. The first stage was a minimum distance classifier with respect to point sets while the second stage utilized the first stage binary valued outputs to implement a Bayes decision. The feature extraction and measure selection problems were examined experimentally. Feature calculation algorithms were developed which are generally applicable to time varying signals. The effectiveness of an algorithm for selection of a set of candidate measures was verified. Experimental results indicated that, for the particular data with which the research was concerned, an assumption of Markoff-1 dependency between sequential first stage decisions of the pattern classifier was warranted. A pattern classifier which was based on this assumption classified 97.9 of the patterns presented to its input correctly.