Machine Learning Forecasts Prognosis of COVID-19 Patients

A machine learning-based risk score can help providers identify COVID-19 patients who are most likely to experience negative outcomes, according to a study published in The Journal of Infectious Diseases.

For more coronavirus updates, visit our resource page, updated twice daily by Xtelligent Healthcare Media.

With the rapid spread of COVID-19, health systems experienced a surge in urgent care visits and hospital admissions. To better predict which patients were at greatest risk for experiencing negative outcomes, researchers at Massachusetts General Hospital (MGH) developed a machine learning tool to predict the prognosis of individual patients with COVID-19.

Experts in neurology, infectious disease, critical care, radiology, pathology, emergency medicine, and machine learning designed the COVID-19 Acuity Score (CoVA) to better identify high-risk patients. The tool is based on input from information on 9,381 adult outpatients seen in MGH’s respiratory illness clinics and emergency department between March 7 and May 2, 2020.

"The large sample size helped ensure that the machine learning model was able to learn which of the many different pieces of data available allow reliable predictions about the course of COVID-19 infection," said M. Brandon Westover, MD, PhD, an investigator in the Department of Neurology and director of Data Science at the MGH McCance Center for Brain Health.

Researchers then tested CoVA in another 2,205 patients seen between May 3 and May 14. In this prospective validation group, 26.1 percent of patients experienced hospitalization, 6.3 percent developed critical illness, and 0.5 percent died, within seven days.

The machine learning tool achieved an area under the curve (AUC) of 0.76 when predicting hospitalization, 0.79 when forecasting critical illness, and 0.93 when predicting death.

The team factored in 30 predictors when creating the model, including demographic information like age and gender, COVID-19 testing status, vital signs, medical history, and chest x-ray results. Among 30 predictors, the top five were age, diastolic blood pressure, blood oxygen saturation, COVID-19 testing status, and respiratory rate.

 "Testing the model prospectively helped us to verify that the CoVA score actually works when it sees 'new' patients, in the real world," said first author Haoqi Sun, PhD, an investigator in the Department of Neurology and a research faculty member in the MGH Clinical Data Animation Center (CDAC).

The tool could potentially help providers rapidly and automatically determine which patients are most likely to develop complications from the virus.

"While several other groups have developed risk scores for complications of COVID-19, ours is unique in being based on such a large patient sample, being prospectively validated, and in being specifically designed for use in the outpatient setting, rather than for patients who are already hospitalized," said Shibani Mukerji, MD, PhD, associate director of MGH's Neuro-Infectious Diseases Unit.

"CoVA is designed so that automated scoring could be incorporated into electronic medical record systems. We hope that it will be useful in case of future COVID-19 surges, when rapid clinical assessments may be critical."

Researchers have increasingly leveraged predictive analytics and artificial intelligence tools to forecast COVID-19 outcomes.

A team from NYU Langone Health recently built an artificial intelligence model to predict which patients will have good outcomes from COVID-19 – information that can be especially helpful during a healthcare crisis.

“Telling a provider that someone might need an ICU bed isn’t useful, because ICU beds are already limited to patients in immediate need of higher levels of care. Predicting future needs wouldn’t change clinical management,” Yindalon Aphinyanaphongs, MD, PhD, assistant professor in the Department of Medicine at NYU Langone Health, told HealthITAnalytics.

“We had to figure out what would be useful for clinicians, and we finally settled down on favorable outcomes with COVID-19. We wanted to build something that would give providers a little bit of confidence that their patients will be okay.”