Oxidative Stress Markers and Prediction of Severity With a Machine Learning Approach in Hospitalized Patients With COVID-19 and Severe Lung Disease: Observational, Retrospective, Single-Center Feasibility Study

Abstract

Background: Serious pulmonary pathologies of infectious, viral, or bacterial origin are accompanied by inflammation and an increase in oxidative stress (OS). In these situations, biological measurements of OS are technically difficult to obtain, and their results are difficult to interpret. OS assays that do not require complex preanalytical methods, as well as machine learning methods for improving interpretation of the results, would be very useful tools for medical and care teams.

Objective: We aimed to identify relevant OS biomarkers associated with the severity of hospitalized patients' condition and identify possible correlations between OS biomarkers and the clinical status of hospitalized patients with COVID-19 and severe lung disease at the time of hospital admission.

Methods: All adult patients hospitalized with COVID-19 at the Infirmerie Protestante (Lyon, France) from February 9, 2022, to May 18, 2022, were included, regardless of the care service they used, during the respiratory infectious COVID-19 epidemic. We collected serous biomarkers from the patients (zinc [Zn], copper [Cu], Cu/Zn ratio, selenium, uric acid, high-sensitivity C-reactive protein [hs-CRP], oxidized low-density lipoprotein, glutathione peroxidase, glutathione reductase, and thiols), as well as demographic variables and comorbidities. A support vector machine (SVM) model was used to predict the severity of the patients' condition based on the collected data as a training set.

Results: A total of 28 patients were included: 8 were asymptomatic at admission (grade 0), 14 had mild to moderate symptoms (grade 1) and 6 had severe to critical symptoms (grade 3). As the first outcome, we found that 3 biomarkers of OS were associated with severity (Zn, Cu/Zn ratio, and thiols), especially between grades 0 and 1 and between grades 0 and 2. As a second outcome, we found that the SVM model could predict the level of severity based on a biological analysis of the level of OS, with only 7% misclassification on the training dataset. As an illustrative example, we simulated 3 different biological profiles (named A, B, and C) and submitted them to the SVM model. Profile B had significantly high Zn, low hs-CRP, a low Cu/Zn ratio, and high thiols, corresponding to grade 0. Profile C had low Zn, low selenium, high oxidized low-density lipoprotein, high glutathione peroxidase, a low Cu/Zn ratio, and low glutathione reductase, corresponding to grade 2.

Conclusions: The level of severity of pulmonary damage in patients hospitalized with COVID-19 was predicted using an SVM model; moderate to severe symptoms in patients were associated with low Zn, low plasma thiol, increased hs-CRP, and an increased Cu/Zn ratio among a panel of 10 biomarkers of OS. Since this panel does not require a complex preanalytical method, it can be used and studied in other pathologies associated with OS, such as infectious pathologies or chronic diseases.

Keywords: COVID-19; ML; SARS-CoV-2; biomarker; coronavirus; hospitalization; infectious; lung; machine learning; oxidative stress; prediction; pulmonary; respiration disorders; respiratory; severity.

Figure 1.. Predicted probabilities of belonging to grade 0, 1, or 2 per biomarker for profile A. Values shown in red indicate the specific values for the profile. P: probability.

Figure 2.. Predicted probabilities of belonging to grade 0, 1, or 2 per biomarker for profile B. Values shown in red indicate the specific values for the profile. P: probability.

Figure 3.. Predicted probabilities of belonging to grade 0, 1, or 2 per biomarker for profile C. Values shown in red indicate the specific values for the profile P: probability.