Analysis of time-frequency features for classification of asthma severity level using computerized wheeze sounds

In asthma patients wheeze sounds are produced due to obstruction in lung sounds. Any medication or management of patients is done according mild, moderate and severe condition of asthma patients. Literature review indicates that analysis and classification of wheeze sounds according to severity levels of asthma patients using time-frequency features in different datasets according to location and phase is required to explore more. The objective of this study is to investigate and classify wheeze sounds according to the severity levels (mild, moderate and severe) of asthma patients using time-frequency features. This study focusses on the self-monitoring and self-management of asthma patients using tidal breathing. Segmented and validated wheeze sounds were collected from the trachea and lower lung base (LLB) of 111 asthmatic patients during tidal breathing. The collected data was split into 9 datasets based on the auscultation location, and/or breath phases. For every segment, the frequency-based, spectral integrated (SI) and integrated power (IP) features were computed. Subsequently, a univariate and multivariate statistical analysis were performed on the features to investigate the significant difference of features in details. Classification was then performed using the ensemble, support vector machine (SVM) and k-nearest neighbor (KNN) methods. In addition, two classification frameworks introduced to identify most effective classification of severity levels. Most of the selected individual features and feature vectors frequency-based, SI, IP observed indicated significant difference (p < 0.05) in majority of datasets. Overall, the best PPV for the mild, moderate and severe samples were found to be 100% (KNN), 92% (SVM) and 94% (ensemble) respectively were obtained with IP features. The μ(SD) values of features have not indicated any specific and continues behavior with respect to severity level in all nine datasets. The findings of research illustrate that the distribution of frequency and spectral energy in the recorded signal varies depending on the auscultation location (trachea and LLB), phase (inspiratory and expiratory) and severity levels (mild, moderate and severe). With the consideration of auscultation location trachea-related datasets produce higher effect size than that of LLB-related datasets. For SI and IP features in most comparisons, the ensemble classifier produced best performance in terms of sensitivity, specificity and positive predictive value (PPV). However, frequency-based features indicated highest performance with the KNN and SVM classifier. Trachea-related datasets samples produced the highest classification performance than all other datasets in all type of combinations. The results of validation also have been found above on average. In future acoustic features, deep learning classification technique and feature optimization can be implemented.