The OXYCOVID study is an observational, retrospective, single-center study based on medical data collected during hospitalization. All adult patients (ie, aged >18 years) hospitalized with COVID-19 at the Infirmerie Protestante (Lyon, France) were included, regardless of the care service they used. Patients were not included if they did not consent to the collection and processing of their personal data, if they were pregnant or breastfeeding, or if they had a protected status (ie, curatorship or other legal protection; deprivation of liberty by judicial or administrative decision). The data were collected from the period from February 9, 2022, to May 18, 2022. The final sample for analysis included 28 patients meeting the inclusion and exclusion criteria. All biological data were recorded in each patient’s digital medical record.

The data were first collected and then sorted specifically for each patient after checking for the absence of errors and for missing data. After verification, no discrepancies were found and the clinical and biological data were judged to be usable. The data are fully anonymized. The data included sociodemographics (sex and age), comorbidities, and risk factors (weight, height, smoking status, hypertension, diabetes, chronic renal disease, severe chronic respiratory disease, autoimmune disease, immune deficiency, and cancer). COVID-19 vaccines were also recorded, as were data related to the current COVID-19 infection (date of the latest polymerase chain reaction test from a nasopharyngeal swab, variant, and date of first symptoms) and the reasons for hospitalization.

At admission, clinical symptoms were recorded, including the presence of fever (temperature >38°C), cough (yes or no), dyspnea (felt by the patient during the initial questioning), myalgia (felt by the patient during the initial questioning), fatigue (felt by the patient during the initial questioning), diarrhea, oxyge saturation (SpO₂), oxygenation system used (nasal oxygen therapy, mask oxygen therapy, or Optiflow system), and pulmonary embolism (determined by a thoracic computed tomography angiogram). Biological data were taken from venous blood samples during hospitalization and included complete blood count; levels of urea, creatinine, C-reactive protein (CRP), troponin, D-dimer, ferritin, lactase dehydrogenase, and prothrombin; activated partial thromboplastin time; levels of total bilirubin, gamma-GT, transaminase, albumin, and interleukin-6; and an OxyCheck panel including the following 10 markers: zinc (Zn), copper (Cu), Cu/Zn ratio, selenium, uric acid, high sensitivity CRP (hs-CRP), oxidized low-density lipoprotein, glutathione peroxidase, glutathione reductase, and thiols.

Patients were triaged at admission into 3 mutually exclusive grades. Grade 0 represented asymptomatic patients. Grade 1 represented patients with mild to moderate clinical signs (SpO₂ >94% in ambient air or with oxygen [maximum 6 L/min], fever, cough, dyspnea, and respiratory rate between 20 and 30 respirations/min). Grade 2 represented the most severe patients, that is, those with severe pneumonia (respiratory rate >30 respirations/min; arterial oxygen saturation <90% despite fraction of inspired oxygen >40% [>6 L/min]) or a rapid increase in oxygen needs (tachycardia [heart rate >120 beats/min]), sweating, agitation, a need for noninvasive ventilatory support with Optiflow (flow rate <60 L/min at 60% fraction of inspired oxygen, ROX index [oxygen saturation/respiration rate index] >3). Grade 2 also included critical patients with severe pneumonia or acute respiratory distress syndrome (index ROX <3 [if the ROX index was <2.5, the patient was intubated]; Optiflow flow rate >60 L/min at 60% fraction of inspired oxygen [if the Optiflow flow rate was >80 L/min at 100% of fraction of inspired oxygen, the patient was intubated]; partial pressure of oxygen in the arterial blood/fraction of inspired oxygen ≤300 mm Hg) and patients with septic shock; acute, life-threatening organ dysfunction related to a known or suspected infection; an altered mental state (delirium or confusion); oliguria; skin mottling; laboratory evidence of coagulopathy, thrombocytopenia, acidosis, or elevated lactic acid; or hyperbilirubinemia.

Baseline characteristics are reported as counts and percentages for categorical variables and medians and IQR for continuous variables. Each patient was assigned only one grade, so the observations are considered independent. We tested for significant outliers among all the biomarkers, determined if distributions followed a normal distribution using the Shapiro-Wilk test, and used the Levene test for homogeneity of variances. Due to the small sample size, the criteria for outlier values were not strictly applied in the ANOVA. Tukey post-hoc tests were used for multiple pairwise comparisons between severity groups when significant differences were observed in the ANOVA. Finally, an SVM was used to predict assignment for each grade, taking into account the entire analysis profile. Statistical analysis was performed with R (version 4.1.1; R Project for Statistical Computing) and the tidyverse, ggpubr, and rstatix libraries. The level of significance was set at P<.05.

Considering that collinearities within our dataset induced by obvious links between biological parameters disrupted traditional methods (eg, logistic regression or the classification method), we used an SVM to predict assignment to each grade for each patient, considering the entire analysis profile. The SVM algorithm is a supervised learning algorithm that uses training data to build a model to predict or classify new observations. SVMs are a machine learning method for nonlinear regression classification [24-26]. Our approach consisted in using the biomarker database as a training dataset for the SVM model (eg, the SVM training dataset). This resulted in a model capable of predicting the probability that a patient would be assigned a specific grade based on the results of a biological analysis, provided that the values did not exceed those used in the initial dataset; that is, the sum of the 3 probabilities was equal to 1. This method resulted in a profile in the form of a graph establishing the link between a combination of biomarkers and the probability of being assigned to the different severity grades. The SVM analysis was performed using JMP Pro (version 17.2; JMP Statistical Discovery).

Participation in this study was voluntary, and all participants received an individual notice at admission to the ICU, which included informed consent (the individual notice is included in Multimedia Appendix 1). This study was based on data available in the medical record and did not require any additional medical visits or examinations; it also did not lead to any changes in the management or treatment of patients. The data were irreversibly deidentified for the statistical analysis. Participants did not receive any compensation. This study was approved by the ethics committee of Infirmerie Protestante of Lyon (CE-23‐01).

A total of 28 patients met the inclusion criteria. The median age was 75 (IQR 68.75-80) years, and the patients were mostly male (n=17, 60.71%; Table 1). At admission, 8 patients were asymptomatic (grade 0), 14 had mild- to moderate-severity illness (grade 1) and 6 had severe to critical symptoms (grade 3). The most common reported symptoms were cough and fever.

Results for the detection of outliers, the normality test, and the homogeneity of variance are provided in Table 2. The OS biomarkers that met the criteria for the ANOVA were Zn (P=.002), Cu/Zn ratio (P=.009), hs-CRP (P=.02), and thiols (P=.004).

Statistical differences were observed between the grade 0 and grade 1 patients and between the grade 0 and grade 2 patients for the 3 OS biomarkers, but the difference was nonsignificant between the grade 1 and 2 patients (Table 3).

Table 4 shows the results obtained from the SVM (Multimedia Appendix 2 provides detailed results). The clinical classification and the SVM prediction differed in only 2 of the 28 cases. One of these cases was a patient who was clinically classified as grade 1; he was symptomatic but did not meet the criteria for intubation. The biological analysis profile was abnormal (low thiols [4.1], low zinc, very high Cu/Zn ratio [2.2]). The SVM model classified this patient as grade 2 with a probability of 70.7%. He remained hospitalized for 1 month with active therapy, made good progress, and was discharged. The other patient was initially classified as grade 2; he was strongly symptomatic with high dependence on high-flow oxygen therapy. The biological analysis profile was abnormal (Cu/Zn ratio=3.57). The SVM model classified this case as grade 1 with a probability of 65.8%, probably due to slightly decreased plasma thiol and zinc levels. The clinical course of this patient quickly became favorable.

With a misclassification rate of only 7%, we estimated that the training of the SVM model was good enough and that it was able to predict the grade of patients with COVID-19 infection based on the results of the initial specific biological assays (upon arrival).

As an illustrative example, we created 3 biological result profiles (named A, B, and C). These profiles were submitted to the SVM model to determine the probability that they belonged to grade 0, 1, or 2 (Table 5). Profile B had a significantly high Zn level, low hs-CRP, a low Cu/Zn ratio, and high thiols and was predicted to be grade 0. At the opposite end, profile C had low Zn, low selenium, high oxidized LDL, high glutathione peroxidase, a low Cu/Zn ratio, and low glutathione reductase, corresponding to grade 2. Figure 1, Figure 2, and Figure 3 illustrate the predicted probabilities per biomarker and per profile for profiles A to C.

Initially, we assumed that antioxidant supplementation for patients in the ICU would reduce the severity of clinical signs and improve their prognosis. The main challenge of this feasibility study was performing a preanalytical and analytical process for measuring OS biomarkers. As a first observational finding, this study identified 3 biomarkers of OS (Zn, Cu/Zn ratio, and thiols) associated with the severity of patients, especially patients with a severity grade of 0 (ie, asymptomatic) or 1 (ie, mild to moderate severity). As a second analytical finding, we found that the SVM model could provide a prediction of the level of severity based on a biological analysis of the level of OS in a relatively limited cohort and during an epidemic of respiratory infectious disease, with only 7% misclassified cases in the training dataset.

The originality of this work is the determination of the grade of clinical severity through the analysis of specific OS biomarkers using a machine learning model. What distinguishes an SVM algorithm from other classical approaches is that when the response is continuous, the fitted models are called a support vector regression (SVR) model. In a typical regression problem, the goal is to fit a model that minimizes the error between a predicted response and the actual response. In an SVR problem, the goal is to fit a model such that the error between a predicted response and the actual response lies in a range of −ε to ε. This allows for more flexible fitting. In JMP Pro, ε is equal to 0.1. The SVR algorithm doubles the data by creating 2 classes, Y + ε and Y − ε. Then, the same algorithm used for the classification problem is also used for the prediction problem (ie, SVR).

Maximization in SVM algorithms is done by solving a quadratic programming problem. In JMP Pro, the algorithm used by the SVM platform is based on the sequential minimal optimization (SMO) algorithm introduced by John Platt in 1998 [27]. Typically, the SVM quadratic programming problem is very large. The SMO algorithm divides the overall quadratic programming problem into a series of smaller problems. The smaller problems are solved analytically rather than numerically, which means that they produce closed-form solutions. Therefore, the SMO algorithm is more efficient than solving the overall quadratic programming problem. While this univariate analysis did not find any differences between grade 1 and grade 2, the trained SVM model estimated the severity grade from the same OS biomarkers; this could have a significant impact in the management of patients at risk of sudden respiratory decompensation. Although the power of the model would benefit from being improved with clinical and biological data from a larger cohort, this approach could be applied to other types of epidemics and other biological assays, as well as to the investigation of other diseases.

We know that moderate and severe forms of COVID-19 are associated with high OS and a low total antioxidant level, which is measured with complex preanalytical and analytical methods [28,29] that are not generally used in all laboratories, have long delays in receiving results, and are not compatible with management in ICUs [30]. Moderate and severe forms of COVID-19 are accompanied by an increase in inflammatory biomarkers, such as CRP, procalcitonin, the neutrophil to lymphocyte ratio, and hemostasis parameters (activated partial thromboplastin time, prothrombin time, D-dimers, and fibrinogens) [29]. The natural correction of low levels of endogenous antioxidants is slow, with a tendency to recover 3 months after hospital discharge [28].

High OS is not only specific to severe disease or infection, but positively influences their severity. A basic infectiology study with a porcine animal model of infection with the avian-type H1N1 European swine influenza virus has shown that prior infection of animals with mycoplasma generates a high level of OS, increasing the severity of influenza and reducing animal performance; host responses could be influenced by diet [31].

The systemic OS status, determined using many biomarkers, was still significantly increased in recovered COVID-19 patients during the recovery phase. Nonhospitalized individuals with COVID-19 presented signs of systemic OS, which is longitudinally associated with the development of post–COVID-19 condition. Regular assessment of plasma thiol levels as a monitoring biomarker and supplementation in cases of deficiency might be potential therapeutic targets at the onset of post–COVID-19 condition [32].

From the beginning of the epidemic, plasma thiol assays at admission were a promising tool to predict ICU admission in patients with COVID-19 [19]. The principles of nutrition for patients in intensive care follow validated recommendations with a high level of evidence [33]. Faced with inflammatory syndrome and significant catabolism, caloric and protein intake must be sufficient to combat muscle wasting and the risk of infection. In products intended for parenteral or enteral nutrition, specific micronutrients or macronutrients intended to modulate inflammatory processes and optimize the body’s immunological or metabolic responses are added, such as the antioxidants β-carotene, vitamin A, vitamin C, or vitamin E, in order to combat OS [34], which plays a central role in the development of major inflammatory states and visceral failure through the production of reactive oxygen species (free radicals), which influence the consumption, distribution, and level of antioxidants [35]. Lowered levels of antioxidants are common and are associated with an increase in morbidity and mortality [36]. Two meta-analyses showed that supplementation with antioxidants (eg, vitamins and trace elements) among ICU patients is associated with a reduction in mortality [20,21] and a reduction in mechanical ventilation time [20], but two more recent meta-analyses showed that the evidence for these benefits remained questionable because they were measured in small study populations and with highly variable cocktails of antioxidant nutrients [23,37]. Currently, it is not recommended to provide massive doses of antioxidants to patients treated in intensive care.

The main limitation of this study is the sample size, with less than 10 patients in the grade 0 and grade 2 groups. Despite the small number of patients, we observed a statistical difference between grade 0 and grade 1 patients in terms of biomarkers associated with severity. The small sample size was caused by a short study inclusion period due to the hospital context and the ongoing crisis, which was not suited to the inclusion of more patients or to the inclusion of other hospitals. The medical and nursing team has continued to collect relevant clinical and biological data concerning patients treated for COVID-19 infection since the end of the study. These new data could be aggregated with those collected during this study and analyzed with an SVM in order to improve the performance of the severity stage prediction. As a beneficial side effect, this study allowed us to test the ability of the clinical research team to study OS biomarkers and identify covariates and risk factors for further investigation after using propensity scoring to adjust for comorbidities and sociodemographic variables of patients.

As early as 2020, applying the SVM method to COVID-19 was reported in the scientific literature, but it was mainly used for diagnosis and prediction of severity in a radiological manner and as a scanner [26]. The prediction of severity stage by SVM has been extended to the analysis of biomarkers classically used in inflammatory and infectious pathologies [27,28] but has not provided a consensus or an infallible score to predict the level of severity. To our knowledge, this study is the first to show the SVM method being used for predicting the severity stage of COVID-19 with pulmonary involvement.

An important point to consider before generalizing the use of prediction scores based on this panel of blood biological OS analyses is the dosage method and standardization of reference values for each analysis and between each laboratory.

In patients with COVID-19 infection, moderate to severe symptoms were correlated with a lowered Zn level, a lowered plasma thiol level, an increased hs-CRP level, and an increased Cu/Zn ratio in a panel of 10 OS biomarkers not requiring complex preanalysis (Zn, Cu, Cu/Zn ratio, selenium, uric acid, hs-CRP, oxidized LDL, glutathione peroxidase, glutathione reductase, and thiols). Applying a machine learning method (an SVM) to the results of this same panel to evaluate the OS level in a retrospective cohort allowed prediction of the level of severity of pulmonary damage in hospitalized patients with COVID-19. 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. Intensive care teams are accustomed to monitoring and managing major inflammatory symptoms, and although it is likely that they need to augment their practices to manage OS and the supply of antioxidants, the biological exploration of OS requires highly complex preanalytical and analytical processes that, to date, are highly limiting factors in the development and dissemination of relevant OS assessments. The fragility and thermolability of certain biomarkers means that any redox reactions that might take place in the sample must be halted; thus, the mandatory use of −80 °C conditions throughout the preanalytical phase makes analysis complex and in some cases even technically unfeasible. We have therefore developed an OS panel with 10 biomarkers that overcomes all these constraints without compromising scientific relevance by performing an assessment of OS based on a series of biomarkers that do not require complex preanalysis (simple freezing at −20 °C); these biomarkers were selected with a machine learning approach in 2009 [23]. The results from this panel showed real significance in estimating OS with robust and reproducible results. Iterative training of this SVM model with a biomarker panel in a larger cohort of patients with OS would give it more power and precision. By determining the ideal weighting of each of the biomarkers with a scientifically validated algorithm, we could obtain a score to estimate the level of OS. A Belgian intensive care team assessed total antioxidant capacity with an electrochemical method (eg, the Total Antioxidant Power technology) rather than by the analysis of biological samples (blood, saliva, or urine), finding that it had potential as a less expensive and much faster alternative to the individual analysis of biomarkers linked to pro-oxidants [30]. This would allow intensive care teams to monitor and respond in real time to systemic OS in their patients.

The World Health Organization has already announced the very probable risk of the appearance of a new epidemic caused by new types of severe acute respiratory syndrome or by the “humanization” of the avian influenza A (H5N1) virus in the years to come [38]. Given the contagiousness and the risk of decompensation, particularly respiratory decompensation, or even death from these viral infections, considering systemic OS in addition to the inflammatory syndrome could be a determining factor in the effectiveness of care in the future. The early and systematic use of machine learning models such as SVMs in clinical, radiological, and biological assessments at the start of an epidemic could provide a better understanding of the functioning and mechanisms of novel pathogens and reduce human loss of life. Sharing of results between different care centers with dedicated databases would accelerate learning and improve the accuracy of these statistical tools.

From a public health perspective, individuals are subject to a range of exposures through the environment (air pollution, solar radiation, and temperature variations) and lifestyle (lack of sleep, psychological stress, tobacco, and poor diet); this is summarized in the term “exposome.” This directly impacts the OS level and therefore the body’s ability to defend itself in the event of infection. Taking action to maintain a healthy environment and lifestyle would help mitigate the severity of diseases such as COVID-19 [39].