Toward AI-Driven Cough Sound Analysis for Respiratory Disease Diagnosis

Respiratory diseases represent a significant health concern that attracts scientists and health professionals due to their impact on public health. Various techniques are used to identify respiratory conditions, from diagnosis to therapy, such as Computer Tomography (CT) scans and Chest X-ray (CXR) images. However, these methods require expert radiologists for manual examination, posing time challenges. This is especially critical during a pandemic, where swift detection and early treatment are paramount. In this regard, an AI-based diagnostic system for automatic identification would play a major role, particularly regarding the analysis of patient cough patterns. Coughs, as a diagnostic cue, provide distinctive information about glottis behavior related to different respiratory pathological cases. This work proposes a comparative investigation of respiratory disease detection techniques using cough sounds, with COVID-19 as a case study. Our study involves the application of three distinct models: a Convolutional Neural Network (CNN) model, a CNN-Support Vector Machine (CNN-SVM) hybrid model, and a transfer learning-based model.We methodically evaluate different model architectures, ranging from custom-built networks to pre-trained deep models, applying spectrogram or Mel-Frequency Cepstral Coefficients (MFCC) in transfer learning-based feature extraction, to determine which is the best approach in terms of accuracy, precision, recall, F1-score, and loss. The experimental findings highlight the superior performance of transfer learning and MFCC feature extraction upon deep learning results using CNN and CNN-SVM with best validation accuracies ranging from 99.55% to 100%. This significant advantage is attributed to (1) the capacity of transfer learning to utilize prior knowledge from related domains, even in data-scarce target tasks, and (2) the proficiency of MFCC in capturing and representing the rich and essential information embedded within cough sounds. Furthermore, we propose an Android-based smartphone application that serves as a noninvasive and real-time prescreening tool. Our system holds significant importance to complement medical efforts containing respiratory diseases and pandemics in infected areas, making it applicable to any respiratory condition or pandemic where cough is a predominant symptom.