Antigen lateral flow devices are frequently used as rapid point-of-care or at-home testing options for COVID-19, increasing convenience and accessibility of testing to a wider population. However, reading the results of a later flow testing device requires a visual interpretation that may vary from person to person and can lead to false negatives, such as when a “faint line” is missed or disregarded. Recently, researchers from the University of Birmingham, Durham University and Oxford University tested the ability of an artificial intelligence app to provide more consistent and accurate results readings, and found that using the app significantly reduced the incidence of false negatives.
The group worked with UK Health Security Agency assisted test centers and healthcare workers conducting self-testing in order to trial the AI app; more than 100,000 images of tests were submitted as part of the study. The tests were interpreted by both the AI algorithm and a panel of four trained clinicians in order to determine to what degree the AI-processed results agreed with those of trained experts. At the assisted test centers, where the results were initially interpreted by a trained operator at the center, the sensitivity of the results was estimated to be 92.08% and the specificity 99.85% when compared to the assessments of the expert reviewers. Using the AI app, the sensitivity and specificity were increased to 97.6% and 99.99%, respectively.
For the self-test results, sensitivity was increased from just 16% to 100%, and specificity increased from 99.15% to 99.4%. While only 4 true positives were identified by test users, 36 true positives were identified by the algorithm and clinician panel among this study subset. This suggests the AI app could be useful in assisting untrained users to accurately interpret their own test results using a smartphone, especially for certain populations such as visually impaired people, children and very elderly people. The algorithm also has the benefit of returning results rapidly, in less than 2 seconds. This trial was published in Cell Reports Medicine.
“The increase in sensitivity and overall accuracy is significant and it shows the potential of this app by reducing the number of false negatives and future infections,” said Professor Camila Caiado, professor of statistics at Durham University and chief statistician on the project. “Crucially, the method can also be easily adapted to the evaluation of other digital readers for lateral flow type devices.”