Optimum Features Computation Using Genetic Algorithm for Wet and Dry Cough Classification

The nature of cough sound has been considered as one of the important diagnostic tools. For example, wet cough in children may represent lower respiratory tract infections. However, cough classification is not an easy task. It cannot be done easily by community health workers. Therefore, an automated method is needed to help them in classifying the types of cough. Several features extraction methods have been proposed for classifying wet/dry cough with different performances. Using all those features have consequences increasing the computational cost. In this work, we develop a method to select the optimum feature set for classifying wet and dry cough in children. We recorded cough sound from thirty children younger than four years diagnosed with respiratory tract infections. Then, sound features such as Mel-frequency cepstral coefficients, energy, non-Gausianity index, zero crossing, linear predictive coding and pitch were extracted. We implemented genetic algorithm to select the optimum features and artificial neural networks to classify wet/dry cough. The results show that our proposed method could reduce around twenty-five percent of the features used in the computation while keeping the accuracy, sensitivity and specificity higher than 96%. The results are much higher compared to the previous studies which involving pediatric subjects. This significant achievement supports the development of in situ respiratory disease screening in distant areas.