Optimized verification of respiratory sounds characteristics utilizing quantile vectors

Socio-demographic and environmental factors contribute to elevated levels of respiratory diseases and continue to represent a globally growing public health burden in light of a limited scope of applicable diagnostic methods. This paper builds on a previously developed system for classification of respiratory condition abnormalities based on Lung Sound Automatic Verification (LSAV). In particular it documents enhanced configuration reflecting a better tradeoff between computational requirements and verification performance. The LSAV system proposed which incorporates GMM (Gaussian Mixed Models) models and quantile vectors is further expanded by incorporating the quantile vector and MFCC (Mel-frequency cepstral coefficients) capacities. Taking into account time-varying stationarity of LS and evaluating sets of varying selective frames from a 400 ms and 300 ms shift down to a 30 ms to 10 ms shift resulted in close to 0% ERR (Equal Error Rate) in verification of LS abnormalities. Utilization of quantile vectors showed added advantages in more complex respiratory sounds classification. The developed approach documents the potential of LSAV with quantiles based classification as a useful scanning tool bypassing substantial subjectivity and inaccuracies of human based auscultation.