Publication

Abstract : The mortality rate due to respiratory diseases is around 44.0 per 100,000 of the paediatric population in recent years and ranks as the 6th cause of death worldwide. Lung auscultation is a cost-effective, non-invasive diagnostic tool in paediatric population where imaging tests are challenging. A public paediatric dataset containing 1,949 respiratory sounds comprised of four classes namely, continuous adventitious sounds (CAS), discontinuous adventitious sounds (DAS), combination, and normal has been used. The Bark Frequency Cepstral Coefficients have been extracted, employing Bark scale filters and discrete cosine transform to de-correlate them. These coefficients are able to perceive frequency resolution similar to human hearing system with better resilience to irrelevant noise components. These are then fed to a less complex model, Extreme Learning Machine algorithm to classify different respiratory sounds. This was tuned by defining the number of hidden neurons, activation functions, and other hyperparameters. The trained model provided the best accuracy of 98.19% with sigmoid activation function and 25 neurons for four classes. The proposed framework was able to achieve a precision of 100%, 95%, 99% and 100% for CAS, DAS, CAS & DAS and normal class. The proposed model assists the pulmonologists in articulating the symptoms in paediatric population.