Using artificial intelligence to improve COVID-19 rapid diagnostic test result interpretation

Abstract

Serological rapid diagnostic tests (RDTs) are widely used across pathologies, often providing users a simple, binary result (positive or negative) in as little as 5 to 20 min. Since the beginning of the COVID-19 pandemic, new RDTs for identifying SARS-CoV-2 have rapidly proliferated. However, these seemingly easy-to-read tests can be highly subjective, and interpretations of the visible "bands" of color that appear (or not) in a test window may vary between users, test models, and brands. We developed and evaluated the accuracy/performance of a smartphone application (xRCovid) that uses machine learning to classify SARS-CoV-2 serological RDT results and reduce reading ambiguities. Across 11 COVID-19 RDT models, the app yielded 99.3% precision compared to reading by eye. Using the app replaces the uncertainty from visual RDT interpretation with a smaller uncertainty of the image classifier, thereby increasing confidence of clinicians and laboratory staff when using RDTs, and creating opportunities for patient self-testing.

Keywords: SARS-CoV-2; machine learning; smartphone application.

Fig. 1.

Image treatment (Top) and convoluted neural network (CNN) (Bottom) used. The image treatment extracts the grayscale image, crops it to the region of interest, normalizes the image, and squeezes it to a smaller dimension for the CNN. The CNN is a simple binary classifier that includes three convolution layers with max pooling and dropout regularization at 0.25, a dense layer with dropout regularization at 0.5 before the final dense layer.