Analysis of Acute Respiratory Distress Syndrome using Machine Learning

Acute respiratory distress syndrome (ARDS) is a common but not well known critical disease condition that is linked to a high death rate. The fact that chest x-rays for ARDS are not always interpreted the same way is a big reason why it is not recognised more often. We wanted to teach a deep convolutional neural network (CNN) how to find signs of ARDS on chest x-rays. CNNs were first trained on 595,506 chest x-rays from two places to find common chest results like opacity and effusion. They were then trained on 8072 chest x-rays that had been marked for ARDS by several doctors using different transfer learning methods. The best CNN was tried on chest x-rays from both an internal and external group. Six doctors, including a chest radiologist and intensive care medicine specialists, looked over a subset of the images. Four hospitals in the US provided chest x-ray statistics. A CNN was able to find ARDS in 1560 chest x-rays from 455 patients with acute hypoxemic respiratory failure, with an area under the receiver operator characteristics curve (AUROC) of 0.92% (95% CI 0.89% to 0.94%). Its AUROC was 0.933 (95% CI 0.888–0.996) for the subset of 413 pictures looked at by at least six doctors, and its sensitivity was 83% (95% CI 74.0% to 91.1%). The AUROC was 0.93% (0.92–0.95%) when images marked as “equivocal” were taken out of the analysis. Chest x-rays can be used to train a CNN to perform as well as a doctor at finding ARDS. More study is needed to see how well these algorithms work for finding ARDS patients in real time so that evidence-based care is followed or to help with current ARDS research.