Deep Learning Multi-Feature Analysis of Cough Audio Signal for Tuberculosis identification

Tuberculosis nowadays still become the diseases with a high mortality rate, and several developing countries are the largest contributors to TB sufferers, the right early treatment, the greater the hope of recovery in sufferers. An easy and efficient early detection system is needed in this case. the experiment we are currently do by using multi-feature of feature extraction methods, in this work we combine several methods to enhancing detecting tuberculosis from the recorded cough signal, the dataset consists of 4 types, specially for tuberculosis and non-tuberculosis we use the Nairobi dataset passive cough dataset (natural cough) 1000 Positive TB and 1000 negative TB and 36 Heavy cough from the Coswara dataset, where from the initial data there is an imbalance in the amount of data, to overcome this problem we use GAN technique to generate 1000 heavy cough data. we combine several features extraction methods including Delta MFCC, Delta-delta MFCC, and Rich Spectral Contrast which are combined for optimal feature representation. A Convolutional Neural Network (CNN) model is employed for classification. The results is 93,4250 for mean accuracy by using 4-fold classification experimental, The results of this experiment provide hope for the development of a tuberculosis detection system that will be further developed so that it is hoped that it will be able to reduce morbidity caused by TB.