Once again, artificial intelligence shows its potential for sifting through massive amounts of medical test data to deliver actionable results, this time with COVID-19 screening in hospitals and emergency departments. We’ve written numerous posts about AI applications in medicine, often for diagnostics. For example, we covered AI assisting with autism spectrum disorder diagnosis at the University of California Davis and Google Health’s success with AI deep learning to improve breast cancer detection.
A group of researchers from Oxford University and Harvard University developed two AI models for COVID-19 early-detection using routinely collected data in hospital emergency departments (EDs) and hospital admissions. Information is available in a research study preprint from medRxiv and bioRxiv. The preprint had not been independently evaluated at the time of publication. We would generally await peer-reviewed study publication before covering it on Health Tech Insider. However, the dramatic results of AI tools that answer extremely urgent medical needs coupled with reports of the deployment to hospitals lead us to write about this AI feat now.
According to the paper, virus-specific testing can take up to 48 hours and has limited sensitivity. The Oxford/Harvard algorithms use blood gas tests and vital signs and can return the results within one hour. The researchers used data from 115,494 ED visits and 72,310 hospital admissions. In testing the algorithms, the researchers found 77.4% sensitivity and 95.7% specificity among all patients who came to the hospital. For patients that were admitted to the hospital, the results were 77.4% sensitivity and 94.8% specificity. Sensitivity refers to the accuracy of positive tests and specificity measures the accuracy of negative test results. When you sift through the stats, the verdict for this study is that if the models say you test positive for the test, 77.4 times the judgment will be correct – leaving 22.6% false positives. If the test comes back negative, however, it has a much higher chance, about 95%, of being accurate.
The AI algorithms aren’t perfect, but they produce significantly accurate results in a much shorter time than conventional testing. The high accuracy level of negative tests is particularly significant when hospital resources worldwide are strained. AI has once again demonstrated its value in analyzing a vast amount of data rapidly to produce a practical result.
