JMIR Publications

ABSTRACT

Background:

Chest radiographs (CXRs) and electronic medical records (EMR), typically obtained early in patients admitted with COVID-19, are key to deciding whether they need mechanical ventilation.

Objective:

The objective of this study is to evaluate the use of a machine learning model to predict the need for intubation using a combination of CXR and EMR data. We included historical data from 2481 hospitalizations at The Mount Sinai Hospital in New York City.

Methods:

We first developed an image classifier by retraining a pre-trained DenseNet model (transfer learning). Then, in the final fusion model, we used a Random Forest (RF) algorithm taking 41 input variables coming from EMR and the probability score from the image classifier.

Results:

At a prediction probability threshold of 0.5, the fusion model provided 82.2% (95% CI 68%-94%) sensitivity, 82.9% (95% CI 79%-87%) specificity, 82.6% (95% CI 78%- 87%) accuracy, and 0.87 (95% CI 0.80-0.94) area under the receiver operating characteristics curve (AUROC) on the test set. Compared to the image classifier alone, which had an AUROC of 0.62 (95% CI 0.50-0.74), the fusion model showed significant improvement.

Conclusions:

We show that we can predict the need for invasive mechanical ventilation in patients with COVID-19 in a robust and automated manner by combining routinely available chest radiography images along with longitudinal clinical data. Such a model may assist risk assessment and optimize clinical decision making in choosing the best care plan during the critical stages of COVID-19.: COVID-19; medical imaging; machine learning; chest x-ray; mechanical ventilation; electronic health records; intubation; decision trees; hybrid model; clinical informatics