Health Life

Electronic health records can be valuable predictor of those likeliest to die from COVID

Credit: Unsplash/CC0 Public Domain

Medical histories of patients collected and stored in electronic health records (EHR) can be rapidly leveraged to predict the probability of death from COVID-19, information that could prove valuable in managing limited therapeutic and preventive resources to combat the devastating virus, researchers from Massachusetts General Hospital (MGH) have found. In a study published in npj Digital Medicine, the team described how artificial intelligence (AI) technology enabled it to identify factors such as age, history of pneumonia, gender, race and comorbidities like diabetes and cancer as predictors of poor outcomes in COVID-19 patients.

“By combining and clinical expertise, we developed a set of models to forecast the most severe COVID-19 outcomes based on past medical records, and to help understand the differences in risk factors across age groups,” says co-lead author Hossein Estiri, Ph.D., an investigator in the Laboratory of Computer Science at MGH and an assistant professor of Medicine at Harvard Medical School (HMS). “Many prior studies have isolated small subsets of EHR data from after the infection, but ours is the first and largest to use entire historical medical records to try to untangle the role of age as the most important risk factor for COVID adverse outcomes.”

The analytics/medical team drew on the COVID-19 “datamart” that had been created by hospital system Mass General Brigham for research, a repository of frequently refreshed longitudinal data on COVID-19 patients from across the system. Using electronic medical records from more than 16,000 such patients, the MGH team applied a computational algorithm—with a human expert in the loop—to identify 46 clinical conditions representing potential for death after a COVID-19 infection. “Despite relying on only previously documented demographics and comorbidities, our models demonstrated performance comparable to more complex prognostic models requiring an assortment of