Model Predicts Severe Disease in Those With COVID-19

Marcia Frellick

September 28, 2020

Editor's note: Find the latest COVID-19 news and guidance in Medscape's Coronavirus Resource Center.

A new prediction model can help clinicians and hospitals discern which patients with COVID-19 are likely to progress to severe disease and how quickly, researchers say.

Brian Garibaldi, MD, associate professor of medicine at the Johns Hopkins University School of Medicine in Baltimore, Maryland, and colleagues developed the COVID-19 Inpatient Risk Calculator with 24 variables known to be linked with COVID-19, such as age, body mass index, underlying conditions, vital signs, and symptom severity at the time of admission.

Data were gathered from the care of 832 consecutive patients with COVID-19 between March 4 and April 24 at five Johns Hopkins hospitals in Maryland and Washington, DC.

Findings were published online September 22 in the Annals of Internal Medicine.

Model Shows Extremes in Risk

The authors say the model can predict likelihood of severe disease (defined as needing high levels of oxygen support or a breathing machine) or death from 5% to 90%, sometimes flagging that a person is 18 times more likely to progress to severe disease than another patient with COVID-19.

The article gives some examples:

"An 81-year-old Black woman with diabetes and hypertension, a BMI of 35 kg/m2, fever, a respiratory rate of 32 breaths/min, a high (C-Reactive Protein) level, and a D-dimer level greater than 1 mg/L has a probability of progressing to severe disease or death of 80%, 92%, and 96% by days 2, 4, and 7, respectively, after admission."

"In contrast," the authors found, a 39-year-old Latinx man with a BMI of 23 kg/m2, no comorbid conditions, and no fever has a probability of progression of 3%, 5%, and 5% by days 2, 4, and 7."

Garibaldi told Medscape Medical News the model has different accuracies at different time points after admission.

"The first two days, it's 85% accurate and then over the first week, it's about 80%," he said.

Uses for the Information

The information can help providers explain to patients and families the likely trajectory of the patient's disease to help set goals of care, Garibaldi said.

For health systems, it's helpful to understand the likelihood of having an uptick in intensive care unit cases, for example, and whether hospitals have the right medications and space.

The research also pointed out how quickly patients can progress to severe disease or death, Garibaldi said.

"The median time to developing severe disease or death in our cohort was just a little over a day. That suggests that for these patients there may be a very limited window for us to do something," Garibaldi said.

Researchers used a precision medicine analytics platform (PMAP) that includes not just age, comorbidities, and demographics but laboratory data, medications, and patient trajectories.

"We really have a good sense of not only what people look like when they arrive at the hospital but what happens to them over the next 7 to 14 days," he said.

There are many models published or in development, but "this is a really well-done one methodologically," COVID-19 risk prediction model developer Michael Kattan, PhD, chair of the Department of Quantitative Health Sciences at Cleveland Clinic in Ohio, told Medscape Medical News.

He pointed to the researchers' use of "area under the cumulative-dynamic receiver-operator characteristic curve" to evaluate the model's ability to discriminate higher- from lower-risk patients.

"That's an elegant way to assess performance and that separates them from the pack," he said.

Additionally, "they used a very modern approach to selecting the predictors as well as tempering their effects," to help increase generalizability, he said.

Generalizability of the data is listed as a limitation in the paper as the research was done at a single institution.

If the model is systematically validated and shown to improve care, providers could eventually see it incorporated into electronic health records, Garibaldi said.

Kattan noted that the researchers' cross-validation by site helps build confidence in the results.

The researchers randomly removed one of the five hospitals and then tested the model built from the other four on that site to compare results. They repeated that check for each site.

"By their interpretation, it worked well at each center when they did it that way," Kattan said.

The one thing Kattan felt was missing was "the calibration performance of the risk calculator," or the correspondence between a predicted probability and the proportion of people who develop the outcome.

As to whether there are any downsides of using prediction models, Garibaldi acknowledged that there's always a danger that relying on an algorithm can lead to missing cases or misjudging cases. That's why the authors emphasize that the model is not meant to replace a physician's expertise, but rather should be used in conjunction with it.

Kattan says the value judgment is whether the information a physician is already using to judge the probability of a patient becoming severely ill with COVID-19 is more effective than a proposed prediction model.

"In this case, it's quite likely that it's not," Kattan said.

Since these data were gathered in the beginning of the pandemic, much has been learned about interventions, symptoms, and spread. Garibaldi said Johns Hopkins has now taken care of more than 3000 COVID-19 patients.

Longitudinal data, knowledge gleaned on how to use mechanical ventilation, and use of therapies will help refine the model, he said.

The study was funded by the Hopkins inHealth and COVID-19 Administrative Supplement for the HHS Region 3 Treatment Center from the Office of the Assistant Secretary for Preparedness and Response.

The authors have declared no relevant financial relationships. Kattan has declared no financial relationships. He is an author on an upcoming paper that proposes a similar predictive model for progression to severe COVID-19.

Ann Int Med. Published online September 22, 2020.

Marcia Frellick is a freelance journalist based in Chicago. She has previously written for the Chicago Tribune, Science News and and was an editor at the Chicago Sun-Times, the Cincinnati Enquirer, and the St. Cloud (Minnesota) Times. Follow her on Twitter at @mfrellick

For more news, follow Medscape on Facebook, Twitter, Instagram, and YouTube.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.