Inflammatory phenotyping predicts clinical outcome in COVID-19

This study was nested within the CoV-19POC study, a trial assessing the clinical impact of molecular POCT in patients within 24 h of presentation to University Hospital Southampton (UHSFT), a tertiary centre located in Southampton, United Kingdom (UK), with suspected COVID-19. Any adult aged 18 years old and over presenting to hospital with suspected COVID-19 and/or an acute respiratory illness was eligible for the study. Full details of the inclusion and exclusion criteria can be found in the master protocol, linked below. The study was approved by the South-Central Hampshire A Research Ethics Committee: REC reference 20/SC/0138, on the 16th March 2020 [9].

Combined nose and throat swabs were obtained from patients at admission to UHSFT and were inactivated by placing them directly in media containing guanidine thiocyanate and detergent. They were then tested for SARS-CoV-2 using the CE marked QIAstat-Dx® Respiratory SARS-CoV Panel (Qiagen™, Manchester, UK), in a dedicated POCT hub embedded within the medical admissions unit. Results were available after 70 min.

Blood was obtained by venesection in serum separating tubes. Samples were left to clot for at least 30 min prior to centrifugation (2000G for 10 min) in line with manufacturers instructions. Serum aliquots were frozen on the day of venesection at − 80 °C in a fully licensed biorepository, and thawed once for later testing. Sera were aliquoted and handled under containment level 2 conditions within a suitable microbiological safety cabinet in line with Public Health England guidance. Serum samples were not inactivated prior to testing.

Demographic and clinical data were collected at enrolment and intervention and outcome data were collected retrospectively from case notes and electronic healthcare records. To capture how unwell patients were at enrolment into the study, the National Early Warning Score 2 (NEWS2), comprising respiratory rate, oxygen saturations, temperature, systolic blood pressure, pulse rate and level of consciousness, was recorded for each patient as a summary of physiological measurements [10]. Laboratory (other than the cytokine readouts) and treatment data, was collected using a real-time data analytics tool (REal-time Analytics for Clinical Trials; digital Experimental Medicines Team, Manchester, UK).

All patients recruited into the main CoV-19POC study who tested positive for SARS-CoV-2 (n = 193) were eligible for serum sampling, but a sample was only collected in 100 patients. The patients who were not tested (n = 93) were either already discharged from hospital by the time of their result, declined testing, or the research team were unable to obtain a sample from venesection. Therefore, in this sub-study, 100 SARS-CoV-2 positive patients had serum processed for cytokine analysis. These results were not known to the treating clinician and therefore did not bias treatment strategy.

Serum from 100 COVID-19 patients were analysed for the following inflammatory cytokines: interleukin (IL)-6, IL-8, IL-1β, IL-33, IL-10, Tumour necrosis factor alpha (TNF), granulocyte-macrophage colony-stimulating factor (GM-CSF) and interferon (IFN) –γ using the Ella™ platform (ProteinSimple, San Jose, California, US) following the manufacturer’s instructions. This cytokine detection system is based on four multiplexed microfluidics enzyme-linked immunosorbent assays (ELISA) run in triplicate within a cartridge. The cytokine panel was chosen based on review of emerging literature on inflammatory responses in COVID-19 [2, 11].

For this assay, 25 μL of serum was combined with 25 μL of assay-specific diluent, to achieve a 1:2 dilution of the original sample. A quality control was run in each cartridge alongside the samples to be tested. Results were scrutinised and any samples with coefficient of variation higher than 10% between the triplicates were re-run. Analytical validation for the lower cytokine ranges was performed using reference cytokine controls for a 7-point standard curve with known quantities, and the Ella™ measurements were comparable above 0.1 pg/mL (data not shown).

The composite end point was admission to the intensive care, the use mechanical ventilation, and/or death. These outcomes were used in a previous observational studies to assess the severity of infectious diseases, such as H7N9 infection [12] and SARS CoV-2 infection [13].

A pragmatic sample size for this sub-study was driven by the access to patient samples and the availability of the Ella™ cytokine analysis platform.

Baseline patient characteristics were summarised using standard descriptive statistics, with number and percentages for binary and categorical outcomes and appropriate measures for continuous outcomes – means and standard deviations for normally distributed variables and medians and lower and upper quartiles for skewed distributions. Cytokine readouts were standardised by calculating a z-score to allow comparison between samples. Spearman’s correlation coefficients were used to quantify the correlations between cytokines and continuous variables. Mann Whitney U tests were used to test whether there were significant differences by sociodemographic factors in cytokine levels. Logistic regression models were used to explore the relationship between the outcome and independent variables. Models are presented for univariate analyses and for multivariable analyses for each cytokine. The covariates in the multivariable models were prespecified and include gender, age, NEWS2 Score, comorbid conditions, treatment medications (e.g. systemic glucocorticoids) and smoking status. Model discrimination is presented using the area under the receiver operator characteristic curve (AUROC). Given the small sample size and the possibility of overfitting in the model, we have used bootstrapped estimates in 1000 bootstrapped samples to estimate the AUROC.

Clinical and laboratory data were collected as part of CoV-19POC study from 100 patients with PCR confirmed SARS-CoV-2 infection who were admitted to UHSFT (Southampton, UK) between 20th March and 29th April 2020. The median time from positive POCT for SARS-CoV-2 to blood sampling was 3 h 12 min (IQR 59 mins to 17 h 18 mins).

Patient clinical characteristics are listed in Table 1. Patients recruited to the study were predominantly male (61%), current non-smokers (82%) with a mean age of 61.55 years. Most were of white, British ethnicity (73.2%) and the most common comorbidities recorded were hypertension (40.6%), chronic respiratory disease (27.7%), diabetes (20.8%) and cardiac disease (19.8%). The median duration of symptoms at admission to hospital was 7 days. Patients recruited to the study were acutely unwell at presentation with a mean NEWS2 score of 5.2 corresponding to a medium risk of deterioration and prompting an urgent doctor review. Markers of inflammation (C-reactive protein, CRP; lactate dehydrogenase, LDH; Ferritin and D-dimer) were raised. In terms of treatment received, 63 (62.4%) received intravenous antibiotics and 10 (9.9%) received systemic glucocorticoids. Of the total cohort (n = 100), 18 (20.7%) patients died, and 44 (43.6%) patients achieved the composite end point.

Older age was significantly associated with death. This was a large effect size with those aged 70 years or over having a univariate odds ratio (OR) of 28 (95% confidence intervals (CI) 5.94, 139.45) of death. In addition, several comorbidities including hypertension, cardiac disease, COPD, diabetes, dementia and malignancy were also associated with death. The only variable significant on univariate analysis for the composite outcome was higher NEWS2 score (OR 1.27, 95% CI 1.07, 1.50). When looking at standard clinical blood tests, neutrophil count at admission, and neutrophil:lymphocyte ratio were both independently predictive of death, with a univariate OR of 1.22 (95% CI 1.05, 1.42) and 1.20 (95% CI 1.08, 1.33) respectively. The results for treatment with steroids as a predictor for poor outcome (univariate OR 6.25, 95% CI 1.48, 26.47) is likely due to steroids being given mainly in the setting of intensive care. Patients who were given steroids were more unwell prior to their treatment rather than treatment resulting in worse outcome. Treatment strategies also had no bearing on cytokine levels as serum was taken at admission prior to escalation of treatment.

Patient serum cytokine levels measured on admission to hospital are listed in Table 2. On both univariate and adjusted models TNF, IL-6, IL-8, IL-1β and IL-33 were significantly associated with the composite end point, with the biggest effect size seen in IL-1β and IL-33 (OR 10.21, 95% CI 1.11, 93.37 and OR 4.78, 95% CI 1.18, 19.30 respectively). Standard demographic information and NEWS2 score was reasonably accurate at predicting worse outcome, with an AUROC of 0.71. This improved to 0.81 with the addition of the COVID-19 blood panel (comprised of neutrophils, lymphocytes, neutrophil:lymphocyte ratio, CRP, LDH, ferritin, D Dimer and high sensitivity (HS) cardiac troponin I. However, addition of the cytokine panel to demographic and NEWS2 score improved the model to a greater degree than COVID-19 blood panel with an AUROC of 0.85. A combination of the demographic and NEWS2 data, COVID-19 panel blood and cytokine levels further improved the predictive capability of the model, giving an AUROC of 0.91 (see Table 2).

When comparing the cytokine levels with patient characteristics, there were no differences in cytokine levels between male and female participants (see Table E1 in the online data supplement). Furthermore, cytokine levels did not seem to differ with age except for TNF, which was significantly elevated in patients aged 70 years or over (p = 0.006) (see Table E2 in the online data supplement). Cytokines IL-6 and TNF showed moderate correlation with CRP (R² 0.45 and 0.43 respectively), with IL-6 also showing a moderate correlation with LDH. Serum IL-8 levels correlated moderately with LDH and D-dimer levels (R² 0.52 and 0.47 respectively) (see Table E3 in the online data supplement).

We assessed whether cytokine levels were predictive of adverse outcome to a greater degree than standard clinical measurements. Looking at the whole cohort, increased age was the greatest predictor for death (See Table 1, OR 28, 95% CI 5.94, 139.4), and in combination with other baseline demographics gave an AUROC for death of 0.93 (see Table E4 in the online data supplement). The addition of individual cytokine data was no more predictive than the use of standard clinical tests (see Table E4 in the online data supplement). Therefore, given increased age is a fixed variable that could not be modified by potential treatment, we evaluated cytokine levels as predictors of outcome in those under 70 years, given this is the cut off for shielding in the UK. Cytokines TNF and IL-33 were independent predictors of the composite end point with an AUROC for adjusted model of 0.84 (p = 0.02) and 0.83 (p = 0.03) respectively, adding a greater degree of prediction than that of the demographic model and NEWS2 score alone (AUROC 0.77, see Table 3). Whilst clinical blood tests were also predictive of poor outcome in combination (AUROC of 0.88), they did not improve prediction of poor outcome individually in the same way as TNF and IL-33, with neutrophil count as the only significant predictor of poor outcome (OR 1.25, 95% CI 1.09, 1.43, see Table 1.). Furthermore, when using all the cytokines combined with NEWS2 and demographic data, the AUROC for the adjusted model improved further to 0.92. The under 70 years subcohort was too small to provide interpretable results when demographic data, NEWS2 and standard clinical COVID-19 blood panel were used in combination with the cytokine results.