Seeing beyond positivity: computer vision approach to decoding BinaxNOW COVID-19 tests and forecasting negative results

The emergence of COVID-19 facilitated the widespread use, regulatory awareness, and societal familiarity of lateral flow assays (LFA) for self-diagnosis. The increased frequency in the use of LFA’s as well as their increased use by the general population amplified some of the major drawbacks of these tests, namely, the reliance on user interpretation of the test result, the qualitative nature of the result itself, and the lower sensitivity and specificity compared to traditional polymerase chain reaction (PCR)-based diagnostic testing. Furthermore, the results of these tests and the interpretation of them by lay users were increasingly used as a public health tool by authorities, where recommendations were made on how members of the public should act in response to self-interpretation of test results. Our aims were to develop an objective image analysis algorithm that can (1) interpret the result of a popular COVID-19 LFA Rapid Antigen Test (RAT) and accurately estimate viral load, and (2) predict how long after a positive test a user will test negative on the RAT, indicating loss of infectiousness and the ability to be around others without fear of spreading infection. Indeed, we found that we could estimate viral load to within a RMSE of 1.44 Cycle Threshold (C T ) of gold standard PCR testing, and that we can predict the result status of a user upon a follow up visit at least 5 days post-symptom onset with a sensitivity of 99%, a specificity of 73% and an accuracy of 94%. The analysis algorithm was finally embedded in a mobile application for ease of use. The development of an objective image analysis algorithm for estimating viral load combined with smartphone apps represents a significant improvement upon LFA RAT technologies being used by millions worldwide to diagnose a variety of epidemiologically significant diseases.