Spectral pattern of cough sounds as an aid to diagnosis in children with common respiratory diseases

Background: Objectives were to identify discriminatory spectral patterns of cough consistent with common respiratory sounds, through machine learning, to assist diagnosis. Methods: Spectral analysis of cough recordings of 50 children each with crackles alone, wheezing alone and absence of crackles and wheezing. Identification of unique features through machine learning by dividing them into training (75%) and testing (25%) datasets. Feature extraction was done using R python and Librosa programming language. Two class classifications of the features extracted from the training dataset were done using classifier models like support vector machine, random forest, K nearest neighbor and classification and regression tree. Identification of the most suited classifier model that could accurately differentiate between the studied respiratory sounds in terms of sensitivity and specificity. Results: Random forest classifier model using Mel-frequency cepstral coefficient (MFCC) gave the best results in differentiating crackles from wheezing with sensitivity and specificity of 66.67% and 66.67%. Classifier model performance improved when augmented with clinical features (Respiratory rate, history of recurrent nebulization and family history of atopy); providing sensitivity and specificity of 83.33% and 91.67%. Conclusions: Cough features extracted and classified by machine learning can be used for non-auscultatory diagnosis of crackles and wheeze. This raises the possibility to develop smart applications for possible use by non-medical personnel to enhance their capability.