Convolutional Neural Networks for the Molecular Detection of COVID-19

Abstract

The ongoing COVID-19 pandemic caused an unprecedented overburning of healthcare systems and still represents a global health issue with the emergence of COVID-19 variants. The relevance of mass testing for COVID-19 in the find-test-trace-isolate-support strategy suggested by the World Health Organization (WHO) is imperative to reduce COVID-19 transmission. Although real-time polymerase chain reaction (RT-PCR) is considered a reference standard for COVID-19 detection, it is an expensive, lengthened, and laborious process, and problems in RNA extraction can reduce the sensitivity. In this context, the Raman spectroscopy analysis in biofluids is a label-free method performing a suitable cost-benefit application for COVID-19 detection. We propose a Convolutional Neural Network (CNN) architecture that processes spectra images generated by the Raman spectrum and returns the COVID-19 diagnosis of the spectrum sample. The predictive performance of the CNN was compared against several other algorithms widely adopted in the literature. The CNN architecture discriminates COVID-19 with Raman spectroscopy of blood samples with 96.8% accuracy, 95.5% sensitivity, and 98.2% of specificity, representing the best results as well as a promising alternative to distinguish samples. Moreover, we also present a model explanation analysis that contributes to clarifying the salient features taken into account by our CNN.