A new artificial intelligence (AI) algorithm that integrates a patient’s CT scan of the lungs and other clinical information could detect and diagnose coronavirus disease 2019 (COVID-19).
The algorithm, published in Nature Medicine, mimics the workflow that a doctor uses to diagnose COVID-19 and it gives a prediction of positive or negative diagnosis. The AI model could also produce separate probabilities of a patient testing positive for COVID-19 based on CT images, clinical data, or a combination of both.
Researchers trained and fine-tuned the algorithm before testing its detection performance, or sensitivity. They then applied the algorithm to 279 patients—half which tested positive and the other half negative for COVID-19. The algorithm had a significantly higher sensitivity of 84% compared to 75% for radiologists evaluating the images and clinical data. More specifically, the algorithm detected 68% of COVID-19 positive cases, whereas radiologists interpreted all of these cases as negative due to the negative appearance of the disease on a CT scan.
The study, partly-funded by NHLBI, suggests that improved detection could keep patients isolated if scans don’t show lung disease when symptoms first present and when COVID-19 symptoms resemble the flu or common cold, that make it difficult to diagnose the disease.