Use of Voluntary Cough Sounds and Deep Learning for Pulmonary Disease Screening in Low-Resource Areas

Abstract

In low-resource areas, pulmonary diseases are often misdiagnosed or underdiagnosed due to a lack of trained clinical staff and diagnostic lab equipment (e.g. spirometry, DLCO). In these settings, traditional methods of pulmonary disease screening often include a lengthy questionnaire (>30 questions) and stethoscope auscultation. Unfortunately, such tools are not appropriate for general practitioner (GP) doctors or community health workers who have little time or experience diagnosing pulmonary disease. We propose a computer-based deep learning algorithm that could enable rapid screening of the most common pulmonary diseases (COPD, Asthma, and respiratory infection (COVID-19)) using voluntary cough sounds alone. Using a dataset of 348 cough recordings, raw cough recordings were segmented into individual coughs and converted to Mel Spectrogram images. We trained two types of models for comparison, binary and multi-class, using transfer learning with VGG19. The resulting Receiver Operating Characteristic (ROC) curves and the Area Under Curve (AUC) accuracy for each model was calculated to evaluate performance. Binary AUC accuracies were 0.73, 0.70, 0.87, and 0.70 for healthy, asthma, COPD, and COVID-19 respectively, while multi-class AUC accuracies were 0.78, 0.67, 0.95, 0.70. This demonstrates good potential for creating a simple low-cost screening tool that is fast to administer. Future versions of the model will use ongoing data collection to expand to more diseases including tuberculosis and pneumonia.