This study aimed to investigate and classify wheeze sound characteristics according to asthma severity levels (mild, moderate and severe) using integrated power (IP) features.

Validated and segmented wheeze sounds were obtained from the lower lung base (LLB) and trachea recordings of 55 asthmatic patients with different severity levels during tidal breathing manoeuvres. From the segments, nine datasets were obtained based on the auscultation location, breath phases and their combination. In this study, IP features were extracted for assessing asthma severity. Subsequently, univariate and multivariate (MANOVA) statistical analyses were separately implemented to analyse behaviour of wheeze sounds according to severity levels. Furthermore, the ensemble (ENS), k-nearest- neighbour (KNN) and support vector machine (SVM) classifiers were applied to classify the asthma severity levels.

The univariate results of this study indicated that the majority of features significantly discriminated (p < 0.05) the severity levels in all the datasets. The MANOVA results yielded significantly (p < 0.05) large effect size in all datasets (including LLB-related) and almost all post hoc results were significant (p < 0.05). A comparison of the performance of classifiers revealed that eight of the nine datasets showed improved performance with the ENS classifier. The Trachea inspiratory (T-Inspir) dataset produced the highest performance. The overall best positive predictive rate (PPR) for the mild, moderate and severe severity levels were 100% (KNN), 92% (SVM) and 94% (ENS) respectively. Analysis related to auscultation locations revealed that tracheal wheeze sounds are more specific and sensitive predictors of asthma severity. Additionally, phase related investigations indicated that expiratory and inspiratory wheeze sounds are equally informative for the classification of asthma severity.