BACKGROUND
Within primary ARDS, SARSCOV2 associated ARDS (CO ARDS) emerged in late 2019 with peak pandemic in the following two years. Recent work in ARDS has focused on phenotyping this heterogenous syndrome with a view to better understanding its pathophysiology.
METHODS AND RESULTS
We conducted a retrospective study on a level four intensive care unit in patients with CO ARDS between April 2020 and February 2021. In total 110 patients were included with mean age 63.2±11.92 (minimum: 26, maximum: 83), 61.2% (68) were male and 25% (17) had severe ARDS. Mortality was 47.3% (52).
Ventilation settings, arterial blood gases and chest x ray (CXR) were assessed at day one of invasive mechanical ventilation and between day two and three. CXR imaging was analyzed by a convolutional neural network (CNN). A binary logistic regression model for CO ARDS mortality prediction was created based on the best performing variables which were age, PaO2/FiO2 ratio (P/F) at day one and day three and CNN extracted CXR features. Early model performance evaluation on test data (23 patients out of the 110) indicated an area under the receiver operating characteristic (ROC) curve of 0.83.
CONCLUSION
Integration of data accessible in all intensive care units can be used for prediction of CO ARDS mortality using evolving P/F ratio and CXR. These may help tailor treatment and plan early discussion for escalation of care and extra corporeal life support. Machine learning algorithms for imaging classification can assess otherwise inaccessible patterns and may become another form of ARDS phenotyping. These features integrated with clinical variables perform better than either alone.