Deep learning methods for classification of respiratory infectious diseases using infrared spectroscopy data

Abstract

In 2019, the SARS-CoV-2 viruses caused the worldwide coronavirus disease (COVID-19) which led to a large number of infected people and also the deaths, meanwhile the global economy suffers a severe downturn due to the spread and adverse effects of COVID-19. To better combat the spread of COVID-19, medical agencies require a faster method to accurately distinguish between positive and negative pharyngeal swab samples for diagnosing coronavirus infections. ATR-FTIR spectroscopy is employed to acquire infrared spectral signals from positive and negative nasopharyngeal swabs, which are then classified using a deep learning model. A new classification method is proposed, based on deep learning models and feature selection and transformation techniques. This method also demonstrates strong performance when tested on FTIR spectral data of positive and negative samples, achieving high mean accuracy, sensitivity, and specificity over 5-fold cross-validation. For the primary spectral data used, the initial CNN model achieves an accuracy of 81.4%, a sensitivity of 71.7%, and a specificity of 90.2%. A significant performance improvement is observed when using the Chi-square-Test-PLS-CNN model, which achieves better results with an accuracy of 92.5%, a sensitivity of 86.5%, and a specificity of 97.7%. To further enhance classification performance, an early-stopping strategy is applied, leading to the best performance achieved by the improved Chi-square-Test-PLS-CNN model, with an accuracy of 95.6%, a sensitivity of 93.0%, and a specificity of 97.9%.