To identify potential prognostic markers of COVID-19, we collected, within the first 48 hours of ICU admission, serum samples from 90 ICU COVID-19 patients (the “Cytokine Cohort”) admitted during the “second wave” of COVID-19 (November, 2020–February, 2021) together with serial serum samples across different timepoints during the course of their ICU admission. PBMCs collected from 14 of these 90 ICU COVID-19 patients (the “Initial Cohort”) were analyzed by mass cytometry (CyTOF) with a training set of 35 monoclonal antibodies ( Supplementary Table 1 ) designed to detect broad shifts in levels of normal PBMC lineages as well as their activation status and the possible mobilization of tissue resident innate immune cells and bone marrow progenitors into peripheral blood. Based on these data we developed a refined, second-generation, 38-antibody CyTOF panel ( Supplementary Table 1 ), which was used on PBMCs collected from a further 28 of the 90 ICU COVID-19 patients (the “Replication Cohort”). Data from the Replication Cohort were used to validate observations from the Initial Cohort on early alterations in immune responses that could effectively differentiate patients likely to recover after a short ICU stay from those who would either die or have prolonged stays in ICU. The ICU admission sera from all patients in the Cytokine Cohort (which includes all patients in the Initial Cohort and the Replication Cohort) were analyzed for levels of four cytokines: IL-1β, IL-6, IL-10 and TNFα ( Figures 1A, B ). PBMC, but not serum samples, were obtained for 8 healthy, age-matched controls for the CyTOF analyses portion of this study.

Clinical and demographic details of all patients and healthy controls are presented in Table 1 and include age, sex, body mass index (BMI), requirement for ventilation during admission and ICU admission levels of serum CRP, blood D-dimer levels, ferritin levels and white blood cell counts along with their differentials. The average age of the ICU patients was 63.5 years with a 2-to-1 bias towards male patients, consistent with previous patient demographic reports linking more severe COVID-19 with older male patients. Table 1 also bins patients into two clinical outcome groups of “Short-Stay” and “Long-Stay/Died” based on the length of time in the ICU and survival: “Short-Stay” patients are classified as those spending < 6 days in the ICU and were survivors, while “Long-Stay/Died” patients are defined as patients who spent 6 or more days in the ICU or died during their stay in ICU ( Figure 1A, C ). While some patients were transferred to the ICU from a general ward rather than being directly admitted, there was no significant difference between the Short-Stay and Long-Stay/Died clinical outcome groups with respect to the mean days spent in a general ward prior to admission to ICU (p = 0.06).

Importantly, we found no significant differences between the two clinical outcome groups with respect to mean age, BMI, blood clotting parameters (D-dimer levels), serum CRP levels or serum ferritin levels ( Figures 1D, E and Table 1 ). At admission, the mean total PMN counts were significantly increased in the Long-Stay/Died group compared to the healthy controls (p < 0.0001) and compared to the Short-Stay group (p = 0.025) ( Figure 1F ). In our separate analyses of just the Initial Cohort and Replication Cohorts, however, differences in PMN counts were not statistically significant and thus we did not consider this measurement as a useful prognostic biomarker of clinical outcomes in the context of smaller cohort numbers. PBMC counts were also not significantly changed between healthy controls and patients or between our two clinical outcome groups ( Figure 1F ). Thus, while these routine clinical tests follow a broad spectrum of parameters including inflammation, coagulopathy, hypo-immunity and autoimmunity, none consistently proves prognostic in identifying patients who would require an extended stay in the ICU or die. Accordingly, we conducted more detailed immunological examinations focused on a single, specific COVID-19-associated process, namely inflammation.

We began by examining serum levels of IL-1β, IL-6, IL-10 and TNFα at ICU admission in all serum samples from our Cytokine Cohort of 90 COVID-19 patients. Strikingly, we found that the mean ICU admission levels of serum IL-10 (p = 0.004) and TNFα (p = 0.0005) were significantly elevated in the Long-Stay/Died group relative to the Short-Stay group ( Figure 1G ). In the Cytokine Cohort, 43% (39/90) of patients had serum ICU admission levels of IL-10 levels ≥ 15pg/ml and 79% (31/39) of these patients fell into the Long-Stay/Died group. Similarly, 42% (38/90) of patients in the Cytokine Cohort had serum ICU admission levels of TNFα ≥ 10pg/ml and 79% (30/38) of these patients were members of the Long-Stay/Died group. Interestingly, serum IL-10 and TNFα only showed a weak correlation with each other (Pearson correlation coefficient R² = 0.12, p = 0.27) suggesting that each may represent a different aspect or chronology of the inflammatory process. In parallel analyses of serum samples from 10 healthy controls (not matched for age or sex) the mean IL-10 level was 1.0 pg/ml (range 0.56 to 1.8) and the mean TNFα level was 2.4 pg/ml (range 1.7 to 3.9) (data not shown). Since the Cytokine Cohort represented a significant number of COVID patients (n = 90), a parallel replication cohort was not constructed.

Given the significant differences in ICU admission levels of TNFα (p = 0.0005) and IL-10 (p = 0.004) between the Short-Stay the Long-Stay/Died clinical outcome groups, we also examined the subsequent mean maximum serum cytokine levels in post admission samples from patients in the two groups and found an even more significant difference between the two groups for both serum TNFα (p < 0.0001) and serum IL-10 (p = 0.0009) ( Figure 1H ). Intriguingly, many patients in the Long-Stay/Died group who demonstrated modest admission levels of serum IL-10 and TNFα subsequently developed high levels during their stay in ICU, further reinforcing the importance of these two cytokines as predictive measures of patient outcomes and monitoring the trajectory of disease.

While admission levels of serum IL-6 were also significantly different between the two clinical outcome groups (p = 0.007) ( Figure 1I and Table S5 ), we excluded this cytokine from further analyses due to the potential confounding effects of anti-IL-6 receptor antibody (tocilizumab) treatments, which has routinely been administered to COVID-19 patients in British Columbia during ICU admission since February 2021 and such treatments could complicate the interpretation of our results. In this study, only six patients in the combined Initial and Replication cohorts received tocilizumab at the time of ICU admission (two patients in the Short-Stay group and four patients from the Long-Stay/Died group). These patients were not significant outliers with respect to the biomarkers of interest presented in this study and thus were not excluded from further analyses. Finally, there were no significant differences between the two clinical outcome groups with respect to mean serum IL-1β levels at ICU admission (p = 0.205) and thus this cytokine was also not analyzed further ( Figure 1I ). In summary, we found that ICU admission levels of serum IL-10 and TNFα were useful and statistically powerful prognostic markers for clinical outcomes in severe COVID-19.

We examined whether parallel CyTOF analyses of peripheral immune cells sampled at the time of ICU admission could reveal additional prognostic biomarkers that identify patients in the Long-Stay/Died group, particularly among those that had serum IL-10 levels <15pg/ml and/or serum TNFα levels <10pg/ml. PBMC samples were available for 42/90 of the Cytokine Cohort patients and these 42 samples were divided into an Initial Cohort (14 samples) and a Replication Cohort (28 samples).

Using a 35-marker CyTOF panel on the Initial Cohort ( Supplemental Figure 1 ) and a 38-marker CyTOF panel on the Replication Cohort ( Figure 2 ), we saw no differences between the Short-Stay patients and Long-Stay/Died patients with respect to major peripheral blood immune populations. The more focused and larger 38-marker CyTOF panel, used to analyze immune cell subsets in the Replication Cohort, permitted clear identification of broad blood cell lineages (B, T, NK and myelomonocytic) as well as major subsets within each cell lineage leading to 41 distinct clusters based on the variable expression of these cell-surface markers ( Figures 2A–C ). A summary of mean absolute counts and p-values can be found in Supplementary Table 2 . The analyses of the Replication Cohort samples and the Initial Cohort samples revealed general lymphopenia in COVID-19 patients relative to healthy controls with respect to both total CD4 T cells and total CD8 T cells but this failed to discriminate between the Short-Stay and Long-Stay/Died patient groups. Additionally, we observed no significant differences in total B cells in COVID-19 patients compared to healthy controls or between the two clinical outcome groups ( Figure 2D ). Although mean total NK cells, MAIT cells, γδ T cells, DC2/3 and pDC were depleted in COVID-19 patients relative to healthy controls these, too, failed to distinguish the Short-Stay group from the Long-Stay/Died group ( Figure 2E ). Finally, while mean total monocytes and stem cell levels were significantly increased in COVID-19 patients relative to healthy controls, neither total monocyte levels nor total stem levels was able, individually, to distinguish the Short-Stay from the Long-Stay/Died patient groups ( Figure 2F ; Table S2 ). In summary, broad immune subset analyses were insufficient to predict COVID-19 patient clinical outcomes with respect to the length of stay in ICU and/or death in either the Initial Cohort or the Replication Cohort. We, therefore, performed more detailed analyses of immune cell subsets within these broad cell categories to identify more subtle potential differences between the two clinical outcome groups that could assist in the prospective identification of Long-Stay/Died patients.

To reveal a larger diversity of specific immune cell subsets we performed more focused cluster analyses on patient PBMC samples from the Initial Cohort and the Replication Cohort after pre-gating for selected major cellular subsets. To restrict the clustering to the myelomonocytic compartment we performed gated analyses on GM-CSFR⁺(CD116⁺) CD19^- CD3^- cells ( Figure 3A ) and restricted clustering to shared marker channels to enable direct comparison between the Initial and the Replication cohorts. A summary of mean absolute cell counts and p-values can be found in Supplementary Table 3 . These analyses did not reveal subsets that separated Short-Stay from Long-Stay/Died patients with respect to absolute cell counts. To focus more specifically on the monocytic subsets as well as to simplify the cluster analyses, after pre-gating on CD116+ CD19- CD3- cells, we restricted the marker channels selected for clustering to a set of 7 markers useful in defining monocytic subsets (CD45, CD14, CD16, CD11c, HLA-DR, CD123, CD56, see Figs. 3B, C). Interestingly, this strategy revealed a CD11c^low classical monocytic subset (CD45⁺ CD116⁺ CD3ϵ^- CD11c^low HLA-DR⁺ CD14⁺ CD16^-/low CD123^-/low) that, in both the Initial and the Replication Cohorts, was consistently enriched in COVID-19 patients relative to healthy controls (Bonferroni-adjusted p = 0.006, Replication Cohort) and was preferentially enriched in the Long-Stay/Died group relative to the Short-Stay group (Bonferroni-adjusted p = 0.076, Replication Cohort) ( Figure 3D ), though the latter did not achieve statistical significance. The potential prognostic value of the CD11c^low classical monocytic marker was restricted to this subset of classical monocytes in both the Initial and Replication Cohorts and did not reflect underlying changes of total classical monocytes which were unchanged in the two clinical outcome groups ( Figure 3E ). Moreover, for both Initial and Replication Cohorts, total intermediate monocytes (CD14⁺ CD16^int) and total non-classical monocytes (CD14^low CD16⁺), as well as observed subpopulations of these types of monocytes, did not prove useful prognostically ( Figure 3F ). Focusing the analyses further on classical monocytes, we found that a three-marker gating strategy was sufficient to identify the CD11c^low classical monocyte population identified by our multi-dimensional analyses (shown here for one representative sample from each group in the Replication Cohort, where intensity is proportional to relative frequency of cells) ( Figure 3G ). Thus, the prognostically useful biomarker of CD11c^low classical monocytes was detectable in two dimensions in both the Initial and the Replication Cohorts using antibodies to a small set of cell-surface markers.

Because lymphopenia has been a consistent feature of severe COVID-19 and T cell subset alterations have been described, we performed similar in-depth analyses of the T cell compartments by gating on CD3⁺ cells prior to clustering. Consistent with our analyses of major subsets in the previous section, we were able to confirm and extend our and other groups’ findings that more subtle T cell subsets are significantly depleted in patients relative to healthy controls including subsets in the CD4⁺, CD8⁺, MAIT and γδ T cell compartments ( Supplementary Figure 2 ). However, while we gained valuable insight into the altered T cell responses in COVID-19 patients, none of these T cell subsets was prognostically useful in separating the Long-Stay/Died patients from the Short-Stay patients.

Since deeper analyses of multiple cytokines and cell subsets at ICU admission revealed significant differences between the Long-Stay/Died and Short-Stay groups, we sought to combine these findings to generate a streamlined prognostic tool that could accurately predict whether a patient, newly admitted to ICU, was likely to have a subsequent longer stay in ICU or die. Although both serum TNFα and serum IL-10 were significantly elevated in Long-Stay/Died patients relative to Short-Stay patients in the Cytokine Cohort, using Pearson analyses, the length of stay in ICU correlated more significantly with serum IL-10 levels (R² = 0.48, p < 0.001) and maximum IL-10 levels (R² = 0.54, p < 0.001) than with serum levels of TNFα (R² = 0.14, p = 0.19) ( Supplementary Figure 3 ). Thus, we proceeded with only serum IL-10 as our cytokine-based pre-screen portion of a stepwise integrated prognostic tool. As the first step, using a cut-off value of 15pg/ml for serum IL-10 (data from the 90-sample Cytokine Cohort), demonstrated a 79% specificity and 55% sensitivity (positive likelihood ratio 2.6) in predicting that a patient newly admitted to ICU would have a longer stay in ICU or die ( Figure 4G ). This prognostic specificity of 79% is somewhat comparable with that seen in the smaller subsets of the Cytokine Cohort, namely the 14-sample Initial Cohort (86%) and the 28-sample Replication Cohort (100%). Similarly, the prognostic sensitivity of 55% in the Cytokine Cohort is somewhat comparable with that observed in the Initial Cohort (86%) and the Replication Cohort (56%) ( Figures 4A–C, G ). The variations in estimates of prognostic sensitivity and specificity between cohorts (the Cytokine Cohort and its two subsets of the Initial Cohort and Replication Cohort), however, likely demonstrate variations that reflect the influences of random patient sampling and, very importantly, cohort size. These results validate serum IL-10 levels as a pre-screen to identify patients likely to die or to experience a long ICU stay.

We then explored the utility of combining serum IL-10 levels (with a cut-off value of 15pg/ml) with levels of CD11c^low classical monocytes (with a cut-off value of 2.7x10⁷/L) as a stepwise integrated diagnostic tool. With this approach, 100% of the Long-Stay/Died patients were correctly identified in the Initial Cohort and 88% of Long-Stay/Died patients were correctly identified in the Replication Cohort, the latter with a specificity of 100% ( Figures 4D, E, G ). These analyses of all 42 patients in the combined Initial and Replication Cohorts (n = 42) allowed us to predict with 91% sensitivity and 91% specificity (positive likelihood ratio 10.1) the clinical outcome of COVID-19 patients newly admitted to the ICU with respect to the likelihood of extended stay or death in the ICU ( Figures 4F, G ). Thus, our results suggest that a simple screen of two biomarkers at the time of ICU admission allows for rapid identification of those patients who are likely to die or require extended ICU care ( Figure 4H ) and has clear implications for patient care and health care delivery.

This study was approved by the University of British Columbia Clinical Research Ethics board (H20-00685) and patient blood was collected at St. Paul’s Hospital and Vancouver General Hospital in Vancouver, BC following informed consent. All COVID-19 patients had a positive nasal or tracheal real time reverse transcription polymerase chain reaction (RT-PCR) SARS-CoV-2 test. To avoid unnecessary virus exposure, patient blood was collected in combination with routine care. Patient samples (n = 90) were collected within 48 hours of ICU admission between November 2020 and February 2021. All patients were equally treated with 6mg/day of Dexamethasone. Patient demographics and clinical information are listed in Table 1 . Patient samples were transported to the main campus of the University of British Columbia for further processing. Healthy control blood samples (n = 8) were collected from age-matched volunteers who showed no COVID-19 symptoms or other illnesses and who had no history of COVID-19.

Combining patients who spent 6 or more days in the ICU with patients who died during their stay in ICU into a single clinical outcome category of “Long-Stay/Died” is based on the not unreasonable assumption that both sub-groups of patients can be defined here as having more severe COVID-19 than those spending < 6 days in the ICU and were survivors. The application and use of these two clinical outcome categories were more powerful in identifying immune differences between categories that the use of more simple categories of “survived versus died” or “did or did not require subsequent ventilation” (data not shown). The choice of 6 days as the cut-off was based upon iterative empirical statistical analyses of immune data: the cohorts were sequentially divided into two test sub-groups (for example, “<1 day” and “1+ days”, “<2 days” and “2+ days”, “<3 days” and “3+ days” etc.) and the immune data of the test subgroups were compared by Student’s t-test to identify the sub-groups with distinct clinical outcomes that had the greatest statistical significance (by p-value) with respect to differences in immune markers.

Blood designated for peripheral blood mononuclear cell (PBMC) analysis was collected into citrate-coated BD Vacutainer™ Glass Mononuclear Cell Preparation (CPT) tubes and PBMCs were isolated within four hours following collection according to manufacturer guidelines. Red blood cell lysis was performed with ACK lysing buffer (Gibco) for 10 minutes. Isolated PBMCs were frozen in fetal bovine serum (Gibco) with 10% dimethyl sulfoxide and stored in liquid nitrogen. Blood designated for serum analysis was collected into BD Vacutainer™ Serum Separation Tubes (SST) and allowed to clot for at least 30 minutes prior to centrifugation at 1200 rcf for 10 minutes and serum collection and storage at -80°C.

Frozen PBMCs were thawed at 37 °C and washed with RPMI 1640 containing 10% FBS and 25U nuclease (Thermo Fisher Scientific, Cat. #88700). Between 1-4 x 10⁶ cells per sample were used for antibody staining. Prior to fixation, all centrifugation steps were performed at 500 rcf and 4°C. All reagent dilutions were prepared according to manufacturer instructions unless stated otherwise. For live/dead cell analysis, PBMCs were stained with Cell-ID™ Intercalator-Rh (Fluidigm, Cat. #201103A) for 15 minutes at 37°C and washed with MaxPar® MCSB. Prior to surface staining, cells were incubated with human TruStain FcX™ (Biolegend, Cat. #422302) for 15 minutes at 4°C and stained with a surface antibody cocktail for 30 minutes at RT (see Supplemental Table 2 for complete list of antibodies). The MR1-5-OP-RU tetramer was incubated together with the antibody cocktail. After incubation, the cells were washed and incubated for 30 minutes at RT with the secondary anti-APC antibody. Prior to fixation and nuclear staining, PBMCs were washed with MaxPar® MCSB (Fluidigm, Cat. #201068) and incubated in MaxPar® Fix and Perm Buffer (Fluidigm Cat. #201067) and Cell-ID™ Intercalator-IR (Fluidigm, Cat. #201192A) for 1 hour. Post fix, all centrifugation steps were performed at 900 rcf and 4°C. To prepare for CyTOF acquisition, PBMCs were washed with MilliQ water and resuspended in EQ™ Four Element Calibration Beads (Fluidigm, Cat. #201078). Samples were acquired with a CyTOF®2 mass cytometer. An average of 400,000 events were collected for each sample at a flow rate of 45µl/min.

Serum cytokines IL-6, IL-10, TNFα and IL-1β were quantified using the Simoa HD-1 platform from Quanterix (Billerica, MA) according to manufacturing guidelines and as specified by Stukas et. al (55).

All data files were normalized (https://github.com/nolanlab/bead-normalization) and events of interest were manually gated with the FlowJo gating software (BD Biosciences). Dimensionality reduction and clustering were performed with Uniform Manifold Approximation and Projection (UMAP) and Rphenograph respectively, as provided in the bioconductor package Cytofkit (https://github.com/JinmiaoChenLab/cytofkit2). The input files were equally down sampled and cytofAsinh was used as the transformation method. The Rphenograph k was set to the default of 30. The dimensionality reduction and clustering were both performed on the entire data set as well as separately on manually pre-gated populations. The assignment of clusters to specific immune cell subsets based on surface marker expression was guided by the cell-type definitions described in Supplementary Table 4 .

Sample size and statistical tests are indicated in figure legends and all graphs and statistical tests were generated using GraphPad Prism (GraphPad Software, La Jolla California, USA). A test was considered statistically significant at a probability of < 5% (p < 0.05) and we did not assume a Gaussian distribution. Where indicated, the p-values for independent t-tests were adjusted for multiple comparisons (Bonferroni). Error bars represent mean ± SD. UMAP plots and heatmaps were exported from Cytofkit and experimental outline figures, including the graphical abstract were created using BioRender.com. Figures were assembled in Microsoft PowerPoint.