Vocal Biomarker Based COVID-19 Detection Using DNN and Transfer Learning ResNet50

The aim of this study is to automate the detection of COVID-19 patients by analysing the acoustic information embedded in cough samples. COVID-19 is a respiratory disease having cough acoustics as a common symptom and indicator. The primary focus is classification of generated deep features from analytical and mathematical representation of cough acoustics using signal processing techniques Mel-frequency cepstral coefficients (MFCCs) and Mel-spectrogram. MFCCs provides feature vector representation of cough signal and is used as an input for deep neural network (DNN) to generate deep features. Transfer Learning ResNet-50 based Convolutional Neural Network (CNN) model is used to generate deep features from image representation of cough in the form of Mel Spectrogram. Dataset labelling is done with two categories of COVID-19 and Non-COVID-19 classes. Among them, we have used 70% of the dataset for training and 30% for testing purposes. The deep features generated from MFCCs and Mel Spectrograms are concatenated along with a feature value output from a DNN having Metadata as input. The final concatenated feature vector is sent for Softmax based classification. By completing the whole process, we obtained the training AUC (Area Under Curve) (ROC) 95.39%, validation AUC as 88.19% and testing AUC as 88.76%. The analysis of final AUC with epoch curve shows constant increase in training AUC and convergence of validation and testing AUC at certain value representing model training as perfectly fit and no overfitting-underfitting problem.