Variational mode decomposition based prediction model for cough sounds using Xception-GRU classifier

Cough is a popular pre-screening tool for the most respiratory disease diagnoses. Discriminating the abnormal cough from normal one by human ear is however challenging as minute pathological changes in auditory pattern go unnoticed and computer assistance is sought for the diagnosis. This study aims at developing a prediction model by analyzing the different frequency components of the signal in multiple dimensions. As a first step, the cough sounds are decomposed using variational mode decomposition (VMD) for five intrinsic mode functions (IMFs) followed by visualizing every IMF in bark frequency scale as spectrogram. The sequentially combined spatial features of all five spectrograms extracted by Xception model when classified using a gated recurrent unit (GRU) based classifier produced a classification accuracy of 97.22%. The fusion strategy adopted in this study has recorded a good performance by the virtue of reduced feature dimensionality. In addition, the better discrimination in the extracted features (low and high frequencies) using VMD, bark scale spectrograms associated with individual's sense of hearing, reduction in the number of parameters using separable convolutions Xception Model and dynamic capturing of sequential features with faster convergence using GRU resulted in identifying the optimal frequency scale of representation for cough sounds and also enhances the overall classification accuracy compared to existing methods in literature.