High-performing age-stratified machine-learning models for the prediction of the severity of COVID-19 progression

By Dr. Liji Thomas, MDReviewed by Danielle Ellis, B.Sc.May 4 2022

Hundreds of millions of cases of coronavirus disease 2019 (COVID-19) have been documented over the last two years during the ongoing pandemic caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). With many of these being severe enough to merit hospitalization, it has been a focus of research to understand how individual patients can be stratified for risk to ensure that the always-limited healthcare resources are allocated to those who truly need them.

A new paper describes an age-related risk model that has been optimized for this purpose.

Introduction

Though most people with COVID-19 have asymptomatic or mild infections, a significant minority develop severe disease requiring hospitalization, oxygen supplementation and/or intensive care unit (ICU) admission. Risk factors for severe disease include increasing age and the presence of chronic medical conditions associated with immunosuppression and/or inflammation. These include cardiovascular risk factors such as hypertension, diabetes, metabolic syndrome, chronic lung disease, and chronic kidney disease, and patients with solid organ transplants or cancers.

Understanding the risk of a given individual is important when selecting the level of care to which that patient is to be assigned when making clinical decisions and for accurate research. Earlier studies have used machine learning to help classify COVID-19 outcomes based on the admission parameters. Further analysis may identify specific risk factors at the individual and population levels.

Both, however, come mostly from studies carried out in higher or specialized centers, show a more frequent occurrence of adverse events than is commonly observed elsewhere, use parameters not typically measured in most clinical settings, and do not analyze the differences in risk factors in hospitalized patients by age group.

The current study, published online in Scientific Reports, aims to use sophisticated machine learning models that stratify patient data by age group to predict disease progression in almost 7,000 patients studied in community hospital settings.

The study occurred over a period of Delta variant predominance but after the mortality rate had gone down with newer care protocols. The researchers used only those tests that are routinely used in all COVID-19 patients.

What did the study show?

The patients in this study had severe outcomes in just over a tenth of the cases. In all cases, the body mass index (BMI), age, and sex were associated with severe disease.

When divided by age into those between 18 and 50 years vs. those above 50, they also found associations between the risk of severe disease in the former group and heart failure and cardiomyopathy. In the latter age group, however, the risk factors for severe disease included dementia and the use of vasopressors within an hour of admission with COVID-19 or the first positive test for the disease.

When the vital signs (blood pressure, heart rate, respiratory rate, oxygen saturation, and body temperature) were aggregated into a single indicator, in addition to the above factors, the resulting classifiers performed well in distinguishing the true positives from the true negatives.

Using the algorithm called the Gradient Boosting Decision Tree (GBDT), the highest true positive rate in the younger subgroup was 75%, comparable to the true negative rate of 74%, with the area under the receiver operator curve (AUROC) being 78%, indicating reasonably good performance.

In the older group, too, this algorithm produced true positive rates above 72%, with a true negative of 73%. The AUROC of 81% again indicates the capability of this approach to predict disease progression with reasonable accuracy.

The same approach also allowed the model’s performance to be assessed across age groups, showing it to perform with an AUROC value of approximately 80% in all groups up to 56 days from the first positive test. Mortality prediction also showed an accuracy of 82% overall, with 79% for the younger and 75% for the older subgroups.

What are the implications?

The risk models used in this study differentiated the risk factors for disease progression in younger vs older patients, allowing, for the first time, age-dependent prediction of severity in adults hospitalized for COVID-19 early in the disease course, at multiple times points. Unlike most earlier models, these findings are based on routinely administered tests and vital signs.

Importantly, the study showed that admission findings, in terms of vital signs and laboratory data, outweighed the presence of underlying illness as a predictor of critical and fatal outcomes from COVID-19. This understanding is probably the result of using GBDT, a sequential ensemble model that removes the effect of age from the predicted risk, thus allowing the effects of other risk factors to be clearly seen.

A history of cancer, cardiomyopathy, and chronic obstructive pulmonary disease (COPD) was more likely to be associated with severe disease in younger patients compared to older, while kidney disease is linked to disease progression in the older group. Age itself is a risk factor for critical and fatal outcomes in younger patients, as is the BMI.

Moreover, the admission findings with respect to the need for oxygen, and the vital signs, were the most predictive of disease progression. In this group, other than age and vital factors, higher creatinine and lower calcium levels were of increased importance.

However, laboratory results were more important in the older patients, especially blood urea, AST, lower calcium, and bicarbonate. Coupled with an aggregate of the respiratory rate, blood pressure, and oxygen saturation, these predicted critical disease the best in this group.

As platelets and lymphocytes drop while creatinine values and neutrophil counts rise, the risk of critical disease went up. Low calcium predicted severity in both age groups but more so in the older group.

“For hospitalized adults, baseline data that is readily available within one hour after hospital admission or a first positive inpatient SARS-CoV-2 test can predict critical illness within one day, and up to 56 days later.” Besides the distinction between the role of specific risk factors in different age groups, the findings showed that the vital signs and laboratory findings were more important than a history of illness, or the sex, in predicting severity.

The results of this age-stratified modeling approach provide advanced understanding of current risk factors for severe COVID-19 outcomes and can help inform care decisions and prioritize next steps for research.”