Researchers at the Laboratory of Computer Science from the Massachusetts General Hospital, Boston, conducted a study on COVID-19 death prediction using routine medical information recorded in electronic medical records.
The study titled: “Predicting COVID-19 mortality with electronic medical records”, published in npj Digital Medicine journal of Nature, aimed to predict death after COVID-19 using only past medical information routinely collected in electronic health records (EHR). In this approach, it was possible to understand the differences between different risk factors according to age groups through the use of generalized linear models (GLMs).
Using computational methods and the authors' clinical experience, groups representing 46 clinical conditions were selected as risk factors for death from COVID-19. They used generalized linear models classified by age to predict the probability of complications or death from the disease before the patients were infected.
They used data from 24,215 patients with a confirmed case of COVID-19 (confirmed polymerase chain reaction (PCR) test) between March 3, 2020 and November 10, 2020. The number was reduced to 16,709 as they excluded patients with at least 1 year of medical history, "i.e., a 1-year time difference between the first and last medical record before the COVID-19-positive PCR test." They collected data from 10 hospitals in the Boston metropolitan area that met the characteristics, one of which was a hospital with more than 3,400 beds.
“Despite only relying on previously documented demographics and comorbidities, our models demonstrated similar performance to other prognostic models that require an assortment of symptoms, laboratory values, and images at the time of diagnosis or during the course of the illness,” the authors explained in the study.
Age was the most important predictor of mortality in patients with COVID-19, as there are also several studies showing that mortality is higher in patients older than 65 years. In addition, sex has also been identified as a risk factor, as both in China and the United States more men were hospitalized than women.
During the creation of the predictive models, they developed one based on age, the general model that was used was divided into three age groups with a variation of 20 years: 45 to 65, 65 to 85 and over 85. 35 variables were identified, including chronic disease, respiratory disease, heart disease, race, sex, among others. In the first group there were 17 characteristics associated with higher mortality, including diabetes with complications; in the second group there were 21 characteristics, most of them related to respiratory system diseases, including smoking; the third group also recorded 17 characteristics.
The authors made the following notes when testing the instrument, on the considerations that should be taken into account when analyzing data from an EHR:
- The history of a medical record does not guarantee that the patient currently has (or maybe ever had) the respective clinical condition. It could be that a disease was resolved over time or never even existed.
- Multiple imputation for missing diagnosis and medication records was not possible. As a result, if a patient did not have any EHR from a disease cluster, we assumed that the history of disease was not present.
- Our models did not fully control for all confounders, which could bias some of the findings.
See the full article at the following link: https://www.nature.com/articles/s41746-021-00383-x
