Deep Learning to Detect Findings of the Acute Respiratory Distress Syndrome on Chest Radiographs
Michael Sjoding
March 9, 2022, Wednesday, 3:00 PM - 4:00 PM EST
The Acute Respiratory Distress Syndrome (ARDS) is a common critical illness syndrome that develops in patients with conditions such as sepsis, pneumonia, or trauma. Bilateral airspace disease on chest imaging is a key criterion in the ARDS definition but a major driver of its lower reliability. We trained a deep convolutional neural network to identify findings of ARDS on chest radiographs by first pre-training the network to identify common descriptive chest findings (e.g. opacity, effusion) and then training it to identify ARDS. The resulting algorithm could identify ARDS findings with performance consistent with or higher than individual physicians. In studying strategies for the ARDS model deployment, we also found that model deferral to clinicians in cases when it had uncertainty in radiograph classification was the optimal strategy. Further research is needed to evaluate the use of this model in real-time healthcare environments or its use in ongoing ARDS research.
Dr. Sjoding's health services research interests concern pulmonary and critical care and hospital medicine, with a focus on inpatient hospital quality measures and their unintended consequences, the epidemiology of critical illness, and methods for critical care delivery. He is also interested in leveraging "Big Data" to enhance critical care research and care, and health service research methods including the use of both clinical and large-scale administrative databases, statistical simulation, causal analysis, and multi-level modeling.