Development of enhanced feature set for improved performance on classification of stridor vs. other respiratory auscultation
Loading...
Author(s)
Ravikiran, Anaghaa
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
For a long time now, the medical field has resorted to subjective methods of detecting breathing and respiratory abnormalities. The primary method of diagnosis is to hear the breathing at various locations using a stethoscope and subjectively analyzing the sounds. Automation of classification of respiratory auscultation has been a significant field of study among various groups of researchers. Auscultation refers to the process of listening to sounds generated by internal organs such as the lungs and heart with the help of a medical device as part of diagnosis. The research presented here aimed to address the need for developing a comprehensive feature set, with minimum number of features, which is intended to have improved performance in distinguishing several classes of pulmonary disorders with scattered data. While many deep learning-based solutions exist, the key is to ensure that the proposed set not only performs well on the trained set of classes but has a competitive performance on newer, unseen classes with just few data samples. The final feature set derived consisted of 18 features, which included 10 MFCC and MFCC Delta features, 3 bark spectral features and 3 spectral features. This feature set was observed to produce an accuracy of 87.62%, 83.57% and 85.15% on asthma vs. COPD, COPD vs. pneumonia, COPD vs. normal respectively, where COPD class served as the control group
Sponsor
Date
2025-04-30
Extent
Resource Type
Text
Resource Subtype
Thesis