Skip to main content

05-06-2021 | Rheumatology | News | Article

EULAR 2021

Machine learning algorithm could predict COVID-19 acute respiratory distress syndrome

Lucy Piper

medwireNews: Researchers have identified a machine learning algorithm that could accurately detect COVID-19 acute respiratory distress syndrome (ARDS) in patients with rheumatic diseases.

“ARDS is a life-threatening complication of COVID-19,” pointed out Zara Izadi (University of California, San Francisco, USA), speaking at the EULAR 2021 Virtual Congress. She noted that little is known about its risk factors among patients with rheumatic diseases, despite them being at a slightly increased risk for poor outcomes.

“Predicting ARDS can guide clinical risk stratification and the treatment of COVID-19,” she said.

Using data from the COVID-19 Global Rheumatology Alliance Provider Registry, the researchers tested seven machine learning algorithms using 75 possible predictors for ARDS, including demographics, medications, and comorbidities.

They trained and tested the algorithms using data on 2592 patients from US healthcare systems, which showed that ARDS was reported in just 5.7% of cases. They then conducted testing using datasets from outside the USA, including Italy, Sweden, Brazil, and Argentina.

Prediction analysis in 826 individuals identified gradient boosting machine (GBM) as the best algorithm among the seven tested. It correctly identified ARDS in 72% of cases, with a specificity and sensitivity of 72%, and an area under the receiver operating characteristic curve of 78%.

By comparison, the remaining algorithms performed worse overall. The Classification and Regression Tree had superior accuracy, at 78%, but sensitivity was poor, at 39%, while k-nearest neighbors and Support Vector Machines had superior sensitivity, at 78% and 75%, respectively, but their accuracy was inferior, at a corresponding 60% and 57%.

The GBM was “reasonably generalizable” to external datasets in Italy and Sweden, Izadi noted, but it was less so for Brazil and Argentina. She suggested that this discordance is likely to be primarily driven by differences in the distribution of key predictors, such as smoking and comorbidities, as well as other factors not measured such as socioeconomic status and access to healthcare services.

Multivariable logistic regression analysis identified the top 10 predictors for ARDS from GBM, which were consistent with previously reported factors, and included increasing age (odds ratio [OR]=1.75 per decade), being a current or former smoker versus nonsmoker (OR=2.11), and comorbidities, notably chronic renal insufficiency or end-stage renal disease (OR=7.35), cancer (OR=5.04), morbid obesity (OR=4.91), obstructive lung disease (OR=4.38), and interstitial lung disease (OR=3.69).

Higher doses of glucocorticoids (per 5 mg increase) and moderate or high disease activity also increased the risk, with odds ratios of 1.30 and 1.78, respectively.

Izadi concluded: “Further studies are needed to evaluate the clinical utility of the prediction models for their potential to guide monoclonal antibody treatment of COVID-19 in high-risk patients with rheumatic disease.”

medwireNews is an independent medical news service provided by Springer Healthcare Ltd. © 2021 Springer Healthcare Ltd, part of the Springer Nature Group

5 June 2021: The coronavirus pandemic is affecting all healthcare professionals across the globe. Medicine Matters’ focus, in this difficult time, is the dissemination of the latest data to support you in your research and clinical practice, based on the scientific literature. We will update the information we provide on the site, as the data are published. However, please refer to your own professional and governmental guidelines for the latest guidance in your own country.

EULAR 2021 Virtual Congress; 2–5 June

See the research in context now

with trial summaries, expert opinion and congress coverage