Background: Upper-airway obstruction of trachea narrowing can lead to life-threating manner. Today, it may diagnosed at advanced stage or remained undiagnosed. The objective of the present project was to develop an algorithm for non-invasive identification of trachea stenosis based on acoustic and mouth pressure signals.
Methods: An experimental system was designed using models of the trachea with four obstruction rates. Signals of five healthy subjects which breathed through the tubes were acquired at four different flow rates. Feature extraction of Short Time Fourier Transform was then obtained and for each breathing cycle, spectrogram was calculated. Multi Scale Pyramids method was applied on each spectrogram, in order to calculate the average energy. The data were classified on three levels: Support Vector Machine (SVM) algorithm between each obstruction pairs at different flow rates for each subject based to his own data, Multi Class SVM classifying between all obstructions rates for each subject based on his own data and the data set of all subjects.
Results: Classification rate of 96% and 87% were obtained for the sound and the pressure signals, respectively, for the first level and 86% and 67%, respectively, were obtained for the second level. However, the last level classification was 73% and 38%, respectively .Nevertheless, no difference in success rate of classification between different flow rates was found.
Conclusion: Classifying obstruction rate can be achieved for individuals and while breathing in moderate rate. Enhancing the data sets with more data sets, features and classifying methods may expand the outcome.