Machine Learning Tool Predicts Respiratory Failure in COVID-19 Patients

The coronavirus is proving to be no match for artificial intelligence technology.

A team at The Feinstein Institutes for Medical Research has developed a machine learning (ML) artificial intelligence (AI)-powered clinical predictive tool that can accurately assess patients for the risk of COVID-19-related respiratory failure within 48 hours.

“Respiratory failure is the leading cause of death among patients with COVID-19 and it is imperative that our clinicians on the frontlines have the most unbiased, informative data to aid their decision making in the emergency department,” said Dr. Theodore Zanos, Ph.D., an assistant professor at The Feinstein Institutes’ Institute of Bioelectronic Medicine. “The clinical tool we propose could identify patients who are at greater risk of deteriorating shortly after being admitted to the hospital and helps show that machine learning and AI can be powerful clinical assets in treating this complex virus.”

The predictive tool was developed to help expedite care for COVID-19 patients upon hospital admission. The research, led by Zanos and Douglas Barnaby, M.D., has been published in the Journal of Medical Internet Research.
 
Scientists collected and analyzed electronic health record (EHR) data from 11,525 patients who were admitted across 13 Northwell Health hospitals, New York’s largest health system, between March 1 and May 11, 2020. Of those patients, 8 percent (933 patients) were placed on invasive mechanic ventilation within 48 hours of admission. Using that data to aid clinicians in predicting which patients may be at risk of respiratory failure, help triage patients and reduce mortality, the team created and compared three ML tools to help score patients upon admission. One of the tools, gradient boosted decision trees – commonly known as XGBoost – had the highest average predictive accuracy at 92 percent.
 
To develop the ML clinical models, scientists used measurements commonly recorded in the emergency department, such as vitals and labs, along with demographics. The model performance was compared against and outperformed the Modified Early Warning Score (MEWS) – a ranking that assesses health risk factors – to calculate the chance of respiratory failure within 48 hours of hospital admission.
 
These models proved accurate in the ability to identify at-risk patients for earlier interventions, including closer patient monitoring, critical care consultation and earlier discussions about goals of care that could ultimately improve patient outcomes. The research team plans to implement the clinical tool within some Northwell Health hospitals.
 
“Data-driven, evidence-based clinical prediction tools can transform frontline medical decision making,” said Kevin J. Tracey, M.D., president and CEO of the Feinstein Institutes. “Dr. Zanos is leading research to understand, validate and develop this strategy for COVID-19 care.”
 
Throughout the COVID-19 crisis, Dr. Zanos and his team have studied ML and AI to develop digital tools to aid clinicians, among them the Northwell COVID-19 Survival (NOCOS) calculator. In July 2020, Dr. Zanos published a perspective on how AI and ML tools could be used specifically for COVID-19. Recently, Dr. Zanos developed a new deep-learning vital sign predictive AI calculator called “Let Sleeping Patients Lie,” which identifies if hospitalized patients can remain asleep overnight, and not be woken up for unnecessary vital sign checks.