Articles

https://doi.org/10.1038/s41591-020-1051-9

An inflammatory cytokine signature predicts

COVID-19 severity and survival
Diane Marie Del Valle1,2,3,14, Seunghee Kim-Schulze1,2,3,4,14, Hsin-Hui Huang5,6,7,14, Noam D. Beckmann8,
Sharon Nirenberg   8,9, Bo Wang   10, Yonit Lavin10, Talia H. Swartz10, Deepu Madduri10,
Aryeh Stock   11, Thomas U. Marron2,3,10, Hui Xie1, Manishkumar Patel1, Kevin Tuballes1,
Oliver Van Oekelen   8, Adeeb Rahman1,2,3,8, Patricia Kovatch   8,9, Judith A. Aberg   10, Eric Schadt8,
Sundar Jagannath10, Madhu Mazumdar5,6,7, Alexander W. Charney   8, Adolfo Firpo-Betancourt11,
Damodara Rao Mendu11, Jeffrey Jhang11, David Reich12, Keith Sigel10, Carlos Cordon-Cardo   11,
Marc Feldmann13, Samir Parekh3,4,10, Miriam Merad1,2,3,4,10 and Sacha Gnjatic   1,2,3,4,10,11 ✉

Several studies have revealed that the hyper-inflammatory response induced by severe acute respiratory syndrome corona-
virus 2 (SARS-CoV-2) is a major cause of disease severity and death. However, predictive biomarkers of pathogenic inflam-
mation to help guide targetable immune pathways are critically lacking. We implemented a rapid multiplex cytokine assay
to measure serum interleukin (IL)-6, IL-8, tumor necrosis factor (TNF)-α and IL-1β in hospitalized patients with coronavirus
disease 2019 (COVID-19) upon admission to the Mount Sinai Health System in New York. Patients (n = 1,484) were followed
up to 41 d after admission (median, 8 d), and clinical information, laboratory test results and patient outcomes were collected.
We found that high serum IL-6, IL-8 and TNF-α levels at the time of hospitalization were strong and independent predictors of
patient survival (P < 0.0001, P = 0.0205 and P = 0.0140, respectively). Notably, when adjusting for disease severity, common
laboratory inflammation markers, hypoxia and other vitals, demographics, and a range of comorbidities, IL-6 and TNF-α serum
levels remained independent and significant predictors of disease severity and death. These findings were validated in a second
cohort of patients (n = 231). We propose that serum IL-6 and TNF-α levels should be considered in the management and treat-
ment of patients with COVID-19 to stratify prospective clinical trials, guide resource allocation and inform therapeutic options.

A
s of late July 2020, COVID-19 disease, caused by SARS-CoV-2 also referred to as cytokine storm, shares similarities with what was
infection, has resulted in more than 15.5 million infections previously seen in patients infected with other severe coronaviruses,
and 634,000 deaths worldwide. A recent study of hospitals in including SARS-CoV and Middle East respiratory syndrome coro-
New York City, at the initial epicenter of the COVID-19 pandemic navirus12, and bears similarities to cytokine release syndrome (CRS)
in the United States, reported that, during March 2020, 21% of observed in patients with cancer treated with chimeric antigen
patients hospitalized with confirmed COVID-19 died1. These find- receptor-modified (CAR) T cells13. Tocilizumab, an IL-6 receptor
ings are aligned with outcomes observed in the Mount Sinai Health inhibitor, is a US Food and Drug Administration (FDA)-approved
System2,3. There are currently no curative or preventive therapies treatment for CRS in patients receiving CAR T cells14. Several
for COVID-19, highlighting the need to enhance current under- single-center studies have used IL-6 inhibitors to treat patients with
standing of SARS-CoV-2 pathogenesis for the rational development COVID-19 with some clinical benefits15 and reported failures14.
of therapeutics. Beyond IL-6, several cytokines have been shown to be elevated in
Recent studies have suggested that, in addition to direct viral CRS and to contribute to tissue damage. TNF-α is important in
damage, uncontrolled inflammation contributes to disease severity nearly all acute inflammatory reactions, acting as an amplifier of
in COVID-19 (refs. 4,5). Consistent with this hypothesis, high levels of inflammation. TNF-α blockade has been used to treat more than
inflammatory markers, including C-reactive protein (CRP), ferritin ten different autoimmune inflammatory diseases, suggesting that
and D-dimer, high neutrophil-to-lymphocyte ratio6–9 and increased this might be a potential therapeutic approach to reduce organ
levels of inflammatory cytokines and chemokines6,8–11 have been damage in patients with COVID-19 (ref. 16). IL-1 is also a highly
observed in patients with severe diseases. Pathogenic inflammation, active pro-inflammatory cytokine, and monotherapy blocking

1
Human Immune Monitoring Center, Icahn School of Medicine at Mount Sinai, New York, NY, USA. 2Precision Immunology Institute, Icahn School of
Medicine at Mount Sinai, New York, NY, USA. 3Tisch Cancer Institute, Icahn School of Medicine at Mount Sinai, New York, NY, USA. 4Oncological Sciences,
Icahn School of Medicine at Mount Sinai, New York, NY, USA. 5Institute for Health Care Delivery Science, Icahn School of Medicine at Mount Sinai, New
York, NY, USA. 6TCI Biostatistics Shared Resource Facility, Icahn School of Medicine at Mount Sinai, New York, NY, USA. 7Population Health Science and
Policy, Icahn School of Medicine at Mount Sinai, New York, NY, USA. 8Genetics and Genomic Sciences, Icahn School of Medicine at Mount Sinai,
New York, NY, USA. 9Department of Scientific Computing, Icahn School of Medicine at Mount Sinai, New York, NY, USA. 10Department of Medicine, Icahn
School of Medicine at Mount Sinai, New York, NY, USA. 11Pathology, Molecular and Cell-Based Medicine, Icahn School of Medicine at Mount Sinai,
New York, NY, USA. 12Anesthesiology, Perioperative and Pain Medicine, Icahn School of Medicine at Mount Sinai, New York, NY, USA. 13Kennedy Institute
of Rheumatology, University of Oxford, Oxford, UK. 14These authors contributed equally: Diane Marie Del Valle, Seunghee Kim-Schulze, Hsin-Hui Huang.
✉e-mail: sacha.gnjatic@mssm.edu

1636 Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine

Nature Medicine Articles
IL-1 activity is used to treat inflammatory diseases, including
Table 1 | Patient characteristics
rheumatoid arthritis and inherited auto-inflammatory syndromes,
such as cryopyrin-associated syndromes, and has led to sustained Patients with Patients without
reduction in disease severity17. IL-8 is a potent pro-inflammatory SARS-Cov-2 confirmed SARS-Cov-2
cytokine playing a key role in the recruitment and activation of by PCR confirmed by PCR
neutrophils during inflammation18, and, given the frequent neutro- Median age (IQR)—years 63 (53–72) 60 (49–73)
philia observed in patients infected with SARS-CoV2, it is possible
Male 787/1,309 (60.1%) 90/167 (53.9%)
that IL-8 contributes to COVID-19 pathophysiology.
To mitigate inflammation caused by SARS-CoV-2, immuno- Race/ethnicity—Hispanic 577/1,268 (45.5%) 62/167 (37.1%)
modulatory agents, including small molecules and monoclonal Race/ethnicity—African 278/1,268 (21.9%) 46/167 (27.5%)
antibodies targeting cytokines, have rapidly been entering into American
clinical trials4, and many such FDA-approved agents are already Race/ethnicity—White 277/1,268 (21.8%) 43/167 (25.7%)
being used routinely in the clinic in an off-label manner. Given the
Race/ethnicity—Asian 73/1,268 (5.8%) 5/167 (3.0%)
significant side effects associated with the use of these agents, there
is an urgent need to identify biomarkers that can accurately pre- Race/ethnicity—Other 63/1,268 (5.0%) 11/167 (6.6%)
dict which patients will deteriorate from an unchecked inflamma- Obesity (BMI ≥30) 465/1,176 (39.5%) 34/149 (22.8%)a
tory response and help guide rational targeted immunomodulatory Comorbidities— 420/1,268 (33.1%) 67/167 (40.1%)
therapeutic strategies. hypertension
In this study, we asked whether inflammatory cytokine levels can
Comorbidities—diabetes 293/1,268 (23.1%) 34/167 (20.4%)
help predict disease course and outcome in patients with COVID-19.
To enhance the relevance of the cytokine assays, we focused on four Comorbidities—CKD 167/1,268 (13.2%) 27/167 (16.2%)
pathogenic cytokines—IL-6, IL-8, TNF-α and IL-1β—with clini- Comorbidities—cancer 147/1,267 (11.6%) 37/166 (22.3%)
cally available or experimental drugs to counteract them. Clinical (active)
specimens were analyzed on the ELLA microfluidics platform (see Comorbidities—atrial 123/1,267 (9.7%) 9/167 (5.4%)a
Methods). We selected this platform owing to the rapid turnaround fibrillation
time of assay results (within 3 h of sample collection), making these
Comorbidities—CHF 69/1,268 (5.4%) 18/167 (10.8%)a
results potentially actionable.
We followed 1,484 patients hospitalized for suspected or con- Comorbidities—asthma 72/1,268 (5.7%) 13/167 (7.8%)
firmed COVID-19 at the Mount Sinai Health System from the day Comorbidities—COPD 44/1,268 (3.5%) 18/167 (10.8%)
of hospitalization to the day of discharge or death. We measured Comorbidities—sleep 50/1,268 (3.9%) 6/167 (3.6%)
serum IL-6, IL-8, TNF-α and IL-1β levels upon admission and cor- apnea
related these results with clinical and laboratory markers of disease
Smoking—current 55/965 (5.7%) 27/143 (18.9%)a
severity and with disease outcome. We found that elevated IL-6
and TNF-α serum levels at presentation were strong predictors of Smoking—history 293/965 (30.4%) 53/143 (37.1%)
disease severity and survival, independently of other standard bio- SARS-CoV-2 PCR detected 1,257/1,257 (100.0%) 0/167 (0.0%)a
marker measurements of laboratory and clinical severity factors. SARS-CoV-2 Ab detected 65/73 (89.0%) 3/5 (60.0%)
These results suggest that multiplex cytokine profiling could be
used to stratify patients and guide resource allocation and prospec- Maximum severity 505/1,168 (43.2%) 99/150 (66.0%)a
scoreb—mild/moderate
tive interventional studies.
Maximum severity 285/1,168 (24.4%) 22/150 (14.7%)a
Results scoreb—severe
Cohort characteristics and cytokine ranges. We obtained labora- Maximum severity 378/1,168 (32.4%) 29/150 (19.3%)a
tory and health information as part of standard clinical care from scoreb—severe with end
1,484 patients with suspected or confirmed SARS-CoV-2 infection organ damage
and hospitalized at the Mount Sinai Health System in New York City SOFA score = 0 309/870 (35.5%) 42/119 (35.3%)
between March 21 and April 28, 2020, under expedited institutional
SOFA score = 1 163/870 (18.7%) 24/119 (20.2%)
review board (IRB) approval. Using an emergency use approval
from the New York State Department of Health, we implemented SOFA score >1 (median 398/870 (45.7%) 53/119 (44.5%)
the ELLA microfluidics soluble analyte test in the clinical laborato- = 4)
ries to measure four inflammatory cytokines known to contribute Acute respiratory distress 199/1,267 (15.7%) 11/167 (6.6%)a
to pathogenic inflammation in CAR T cell-associated CRS—IL-6, syndrome
IL-8, TNF-α and IL-1β—and assessed their correlation with severity Died within follow-up 269/1,317 (20.4%) 13/167 (7.8%)a
and survival. Of the patients tested, 1,257 had a documented posi- period
tive or presumptive positive SARS-CoV-2 polymerase chain reac- a
Significantly different between PCR-positive and PCR-negative subsets by Fisher’s exact test or
tion (PCR) test, whereas the remaining 167 could not be confirmed. chi-squared test. bSeverity score, described in Methods, is a composite measurement for COVID-
A total of 1,953 specimens were analyzed to quantify circulat- 19 established by Mount Sinai infectious disease and pulmonology teams based on criteria used
ing IL-6, IL-8, TNF-α and IL-1β serum levels using the ELLA rapid in the literature and includes CrCl, ALT levels and use and type of respirators/ventilators and
vasopressors (see Methods). Severity score indicated here is the maximum achieved throughout
detection enzyme-linked immunosorbent assay (ELISA) microflu- the observation period, but, for all other analyses, the severity score at the time of the first
idics platform (Methods and Extended Data Fig. 1a–e). In most of measurement was used. Denominator indicates number of patients with available information.
the 1,484 patients accrued, samples were collected once, typically CKD, chronic kidney disease; CHF, congestive heart failure; COPD, chronic obstructive pulmonary
disease; Ab, antibody.
upon admission to the hospital (median, 1.2 d; interquartile range
(IQR), 0.7–3.0 d). A subset of patients (n = 244) had cytokine mea-
surements performed more than once after admission, although, for still in hospital, whichever was latest) was 8 d (IQR, 3.1–16.0 d, up
all prognostic analyses, only the first available test was used. For to 41 d). Patient characteristics are listed in Table 1. As references,
the entire cohort, the median time available from first cytokine test and to serve as controls, cytokine measurements collected before
to last follow-up (that is, date of discharge, date of death or date the launch of this study were performed in healthy donors and in

Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine 1637

Articles Nature Medicine

**** **** **** correlations with demographics and comorbidities. We hypoth-

**** * ****
**** ******** * **** NS ** **** * NS esized that cytokines are elevated in patients with COVID-19
32,768 compared to healthy donors and non-CRS CAR-T-treated patients
owing to SARS-CoV-2 infection. Of the 1,484 patients hospitalized
Healthy donors with COVID-19 symptoms, 11.7% tested negative for SARS-CoV-2
1,024 CAR T no CRS by PCR and, therefore, were excluded from further univariate anal-
Cytokines (pg ml–1)

CAR T CRS yses. It should be noted that there might have been false-negative
COVID-19
tests for SARS-CoV-2 viral detection based on subsequent tests
32 demonstrating antibodies to SARS-CoV-2 S spike protein in three of
five patients who tested negative for SARS-CoV-2 by PCR. Despite
similar comorbidities, cytokine levels in this subset of patients were
1 significantly lower compared to patients who tested positive for
SARS-CoV-2 infection (Fig. 2a). Of the remaining patients who
tested positive for SARS-CoV-2 by PCR, 1,097 had complete infor-
mation for demographics and comorbidities.
Men had significantly higher levels of IL-6 than women
o D

o D
C S

C S

TN o D

C D
IL IL D

F- F D

C RS
IL -6 S

IL -8 S

IL L- ID

D
TN F-α RS
-6 CR

-8 CR
IL CR

IL CR
N H

N H

N H

β H
VI

TN T VI

VI
V
Fα C
C
O

O
-6 -6

-8 -8

α α

-1 1β

(P < 0.0001), but no sex differences were observed for the other

IL IL

three cytokines (Fig. 2b). With increased age brackets (<50, 50–70
IL-6 IL-8 TNF-α IL-1β and >70 years old), levels of IL-6, IL-8 and TNF-α increased
(Fig. 2b), and the same was observed for age when assessed as a
Fig. 1 | Range of measured cytokines. Detection range of cytokines in continuous variable. There was no association of any cytokine mea-
all tested serum samples from patients with COVID-19 hospitalized at sured with body mass index (BMI). Smoking and race/ethnicity
the Mount Sinai Health System (orange, n = 1,959), in comparison with showed weak but significant univariate associations with IL-6, IL-1β
serum samples from healthy donors (black, n = 9) and plasma samples and/or TNF-α, which were not confirmed after adjusting for the
from patients with multiple myeloma prior to (blue, n = 151) and during other covariates, except for IL-1β and TNF-α, which remained sig-
(red, n = 121) CRS induced by CAR T cell therapy. Heavy bars indicate nificantly higher when comparing Hispanics to African Americans.
median, and error bars represent 95% CI, each value indicated by a dot. We then assessed whether cytokine levels were associated with
Pairwise comparisons by the two-sided Mann–Whitney t-test show comorbidities listed in Table 1. We found that TNF-α and IL-8
significantly higher levels of IL-6, IL-8 and TNF-α in COVID-19 samples were significantly increased in patients with chronic kidney dis-
compared to samples from healthy donors of patients with non-CRS cancer ease (CKD), diabetes and hypertension, whereas TNF-α was also
(****P < 0.0001, ***P < 0.001, **P < 0.01 and *P < 0.05; NS, not significant). increased in patients with congestive heart failure (CHF), based on
Median, mean and range are shown in Extended Data Fig. 1d (error band univariate analyses. IL-6 and IL-8 were elevated in patients with a
indicates the median with 95% CI). HD, hemodialysis. history of atrial fibrillation. No associations were found between
cytokines and active cancer, asthma, chronic obstructive pulmo-
nary disease (COPD), human immunodeficiency virus (HIV) and
patients with cancer who either developed or did not develop CRS sleep apnea.
after CAR T cell therapies19,20. Using multivariable regression models, we confirmed that CKD
We found that IL-6 (P < 0.0001), IL-8 (P < 0.0001) and TNF-α was the only comorbidity significantly associated with elevated cyto-
(P <  0.0001) were significantly elevated in COVID-19 serum kine levels, whereas elevated TNF-α in patients with diabetes and
compared to healthy donor serum or plasma isolated from CAR hypertension were explained by other variables. Of demographic
T cell-treated patients with no CRS (Fig. 1). The four cytokines variables, age and sex (for IL-6) remained significantly associated
assessed had different detection ranges, with IL-6 having the most with cytokine levels as seen in univariate analyses. Therefore, we
dynamic profile, followed by IL-8 and TNF-α (Fig. 1 and Extended included demographics and comorbidities as confounding variables
Data Fig. 1d). In line with previous reports, IL-1β levels were mostly in subsequent analyses. Cytokine levels, as measured by ELLA, were
low or at the limit of detection of 0.1 pg ml−1, even though the assay not significantly affected by timing of testing in relation to hospital
was able to detect various levels of recombinant control cytokines admission. Therefore, this time difference was not considered as a
(Extended Data Fig. 1b). The vast majority of patients, therefore, potential confounder.
presented with elevated cytokines or cytokine storm, but, in con-
trast to the coordinated increase in cytokines during CAR T CRS Association between cytokines and risk of death. Next, we con-
(average Spearman’s r = 0.6), cytokine levels were not as highly cor- sidered factors affecting survival defined as time to death and cen-
related with each other in COVID-19 samples (average Spearman’s sored regardless of cytokines in the overall cohort with univariate
r = 0.4), suggesting differential patterns of cytokine expression and Kaplan–Meier analyses. We found that only age and CKD were sig-
potentially distinct clinical presentations based on the relative pro- nificantly associated with increased risk of death from COVID-19.
file of each independent cytokine (Extended Data Fig. 1e,f). Because We evaluated whether cytokines could distinguish patients based
more than 70% of samples analyzed for each cytokine in COVID-19 on overall survival and disease severity after COVID-19 hospitaliza-
fell within the CRS range based on our post-CAR-T-defined cut- tion. Stratifying patients by cytokine levels of high versus low using
offs, and because we did not have an established cutoff for IL-1β, we the cutoffs described in the statistical analysis section, we found that
decided to separate high versus low values using a cutoff above the each cytokine could predict the overall survival of patients, based on
median for each cytokine in patients with COVID-19. After empiri- the first available measurement after hospital admission. Each cyto-
cal testing as described in the Methods, the cutoffs chosen for fur- kine was independently predictive of overall survival, after adjust-
ther statistical analyses were more than 70 pg ml−1 for IL-6, more ing for demographics and comorbidities—that is, sex, age, race/
than 50 pg ml−1 for IL-8, more than 35 pg ml−1 for TNF-α and more ethnicity, smoking, CKD, hypertension, asthma and CHF (Fig. 3).
than 0.5 pg ml−1 for IL-1β. When considering all cytokines together in the model, all but
IL-1β remained significant, even after adjustment for demograph-
Association with demographics and comorbidities. We used the ics and comorbidities (n = 1,097). This confirmed the relative
first available cytokine measurement in each patient to measure independence of each cytokine tested, with only age (50–70 versus

1638 Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine

Nature Medicine Articles
a
IL-6 IL-8 TNF-α IL-1β

CoV-2 Neg
**** **** **** NS
PCR Pos

16 32 64 128 256 512 16 32 64 128 256 8 16 32 64 128 0.06 0.25 1 4 16

pg ml–1

b
Sex Female
NS NS
Male
**** NS

Age <50
** **** ** NS
50–70 **** **** **** NS
>70 *** **** **** NS

Body mass <30

index NS NS NS NS
≥30

Afr. Amer.
* NS * **
Hispanic NS NS NS NS
Caucasian
*** NS NS
*

16 32 64 128 256 512 16 32 64 128 256 8 16 32 64 128 0.06 0.25 1 4 16

pg ml–1

c
Smoking N
NS NS
Quit NS NS
* NS
NS
NS
NS NS NS NS
Y

Diabetes N
Y
NS * **** NS

Hypertension N
NS **
Y **** NS

N
Chronic kidney NS
disease Y **** **** *

Chronic heart N
failure NS NS
Y * ***

Asthma N
NS NS NS NS
Y

Atrial N
fibrillation
Y **** **** NS NS

16 32 64 128 256 512 16 32 64 128 256 8 16 32 64 128 0.06 0.25 1 4 16

pg ml–1

Fig. 2 | Cytokine levels by PCR status, demographics and comorbidity. Cytokine levels observed in relation to a, SARS-CoV-2 PCR status (negative
indicates patients with COVID-19-like respiratory symptoms with a negative SARS-CoV-2 PCR test) (n = 1,422 independent patient samples);
b, demographics (excluding PCR negative, with data available for sex, age, BMI and race/ethnicity for 1,298, 1,307, 1,174 and 1,131 patients, respectively);
and c, comorbidities (excluding PCR negative, data available for smoking and comorbidity diagnoses, respectively, for 964 and 1,266 individual patients).
Scatter plots indicating individual measurements (dots); thick line is median; error bars representing 95% CI; and statistical analyses by two-sided Mann–
Whitney univariate t-test (****P < 0.0001, ***P < 0.001, **P < 0.01 and *P < 0.05; NS, not significant). Not shown here are COPD, HIV, sleep apnea and
active cancer, which did not show any significant difference for cytokine levels. In yellow highlights are the statistical values that were still significant after
adjustment of all demographic and comorbidity variables, with shade of yellow indicating adjusted P value (light: *, mid: **, high: *** and saturated: ****).
Gray area indicates cytokine levels below the respective cutoff.

<70 years, hazard ratio (HR) = 2.09 (1.25–3.49); >70 versus <50 respectively). Internal validation for this model achieved an uncor-
years, HR = 3.76 (2.24–6.33)), IL-6 (HR = 2.23 (1.61–3.09)), IL-8 rected concordance index of 0.738, a ten-fold coefficient of varia-
(HR = 1.41 (1.05–1.89)) and TNF-α (HR = 1.50 (1.09–2.07)) tion (CV) concordance index of 0.705 and a bootstrap-corrected
remaining significantly associated with decreased survival after concordance index of 0.716. As additional validation, we also
adjustments (P = 0.0049, P < 0.0001, P = 0.0205 and P = 0.0140, performed this analysis using a competing risk model, in which

Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine 1639

Articles Nature Medicine

IL-6 IL-8 TNF-α IL-1β

1.00 1.00 1.00 1.00

Survival probability
0.75 0.75 0.75 0.75
High
0.50 0.50 0.50 0.50
Low
0.25 0.25 0.25 0.25
P < 0.0001, HR = 2.47 0.00 P < 0.0001, HR = 1.84 0.00 P = 0.0001, HR = 1.80 0.00 P = 0.0621, HR = 1.28
0.00
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
Days from ELLA cytokine test to last follow-up

Fig. 3 | Cytokine levels and survival. Survival curves based on each cytokine measured, after multiple variable adjustments for sex, age, race/ethnicity,
smoking, CKD, hypertension, asthma and CHF (n = 1,246). Cox regression model showing overall survival with CIs for each cytokine based on time from
ELLA cytokine test to last follow-up date (discharge, death or still in hospital, whichever comes last), with significance indicated by P value and HR. There
was worse survival if cytokines were high (red, above cutoffs of 70 pg ml−1 for IL-6, 50 pg ml−1 for IL-8, 35 pg ml−1 for TNF-α and 0.5 pg ml−1 for IL-1β) versus
low (blue, below cutoffs). Each line indicates the predicted survival probability over follow-up time, with the error band indicating the corresponding
two-sided 95% CI.

patients discharged alive were considered competing events, and this model achieved an uncorrected concordance index of 0.794 and
patients in hospital were censored, and found the same conclusion, a corrected index of 0.764 (ten-fold CV and bootstrap). In a subset
where high IL-6, IL-8 and TNF-α remained significantly associated of patients (n = 663), Sequential Organ Failure Assessment (SOFA)
with worse outcome regardless of demographics and comorbidities severity scale scores were also available, and we confirmed that IL-6
(Supplementary Table 1). We used the competing risk model in the (HR = 2.9, P < 0.0001), IL-8 (HR = 1.6, P = 0.04) and TNF-α (HR
next analysis. = 1.6, P = 0.03) were associated with poor survival, after adjusting
for all most informative variables above, including increased SOFA
Using cytokines to complement risk stratification. Next, we asked severity (treated either as a continuous variable or as SOFA score of
whether cytokines were of value for risk stratification and survival, ≤1 versus >1).
independent of known laboratory and clinical severity metrics (that Finally, we applied the survival models from our analysis (the
is, temperature, O2 saturation, respiratory rate and severity score primary model: cytokines, demographics and comorbidities; the
as defined in Methods and Extended Data Fig. 2). We first tested secondary model: the primary model plus the markers of inflamma-
whether the four tested cytokine levels were associated with known tion, renal function, myocardial strain and respiratory distress) to
inflammation markers CRP, D-dimer and ferritin and found strong an independent validation cohort of 231 hospitalized patients who
correlations in all cytokines with each measurement, with IL-6 and tested positive for SARS-CoV-2 by PCR collected between April 22
IL-1β additionally associated with fever (Fig. 4a). In addition, IL-6 and June 16, 2020, with available cytokine, demographics, comor-
and IL-8 levels were closely correlated with severity scale (moderate, bidity and laboratory data. The area under the receiver operating
severe and severe with end organ damage), which takes into account characteristic curve (AUC) plots showed that the primary model
lung imaging, creatinine clearance (CrCl), vasoactives and use of performs well between days 3 and 31, during which the AUC ranged
ventilation, whereas TNF-α did not distinguish moderate versus from 0.65 to 0.76 (Extended Data Fig. 4a). The secondary model
severe COVID-19 presentation or use of mechanical ventilation, had somewhat higher AUC, ranging from 0.70 to 0.88 (Extended
but, instead, was only increased with end organ damage. Looking Data Fig. 4d). The integrated AUCs of the two models were 0.68 and
at the predictive value of cytokines on survival after adjusting for 0.74, respectively. The actual and the predicted survival probabili-
levels of CRP, D-dimer, ferritin and all comorbidities, IL-6 and IL-8 ties were similar until day 20, after which the two curves separated
remained independently predictive of survival, therefore showing (Extended Data Fig. 4b,e). The distributions of the prognostic indi-
additive value to these known markers (Supplementary Table 2). ces were not significantly different between the original and valida-
When including additional severity metrics, including severity tion cohorts for the primary model (P = 0.11) and the secondary
scoring, IL-8 was no longer predictive of survival, likely because model (P = 0.06) (Extended Data Fig. 4c,f).
these added parameters were stronger factors for the competing risk Therefore, we conclude that IL-6 and TNF-α are independently
model (Supplementary Table 3). predictive of patient outcomes in terms of both disease severity and
We then investigated correlations of cytokines with an additional survival (Supplementary Table 4). Even after stratifying for risk
series of well-established markers of inflammation, renal function, factors with the strongest P value—that is, severity score, O2 satu-
myocardial strain and respiratory distress for their effect on survival ration and age—IL-6 and TNF-α remained independently predic-
within this cohort. Using unsupervised analyses, neutrophils, white tive of survival, with IL-8 also reaching significance (Fig. 4c and
blood cells, CRP, ferritin, D-dimer, lactate dehydrogenase and low Supplementary Table 5).
O2 saturation co-clustered with all cytokines except TNF-α, which
was more closely correlated with markers of tissue damage such as Effect of medication and treatment on cytokine levels. Although
creatinine (Extended Data Fig. 3). Selecting the most informative our data do not demonstrate a causative role for IL-6 and TNF-α in
variables using a backward elimination process to define each of disease outcome, we wanted to shed light on the effects of various
the available measurements to be used as confounding factors in a treatments on measured cytokines as potential mitigation strate-
competing risk regression analysis of survival along with the cyto- gies should there be a pathogenic effect from these inflammatory
kines, we found severity score, O2 saturation, platelets, low albumin, agents. From a subset of 244 patients with more than one ELLA
systolic blood pressure, D-dimer, albumin, calcium, chloride and cytokine assay performed, and by mapping time from treatment
platelet count remaining. Remarkably, even when using these mea- start to first ELLA test, we were able to assess the effects of various
surements as variables to adjust when assessing the predictive value treatments and experimental drugs on cytokine levels (Fig. 5a). Our
of cytokines on survival in the competing risk regression analysis analysis of a subset of patients with progressive respiratory failure
(n = 802), we found that IL-6 and TNF-α remained significantly and marked systemic inflammation who received off-label treat-
associated with a worse prognosis (Fig. 4b). Internal validation with ment with the anti-IL-6 receptor monoclonal antibody tocilizumab

1640 Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine

Nature Medicine Articles
a IL-6 IL-8 TNF-α IL-1β
Fever <100.4
**** NS NS
°F >100.4 ****
O2 >95
Saturation 90–95 **** * NS NS
**** **** NS **
**** **** NS
% <90 *
CRP <100
mg L–1 >100
**** **** ** ****
D-dimer <1
mg L–1 >1 **** **** **** **

Ferritin <1,000
µg L–1 >1,000
**** **** **** *
Platelets High
NS NS NS NS
Normal **
** NS NS NS
NS
NS *
Low

Severity 1
**** **** **** NS NS
Score 2 **** **** ****
(At first visit) 3 **** **** **** ****
16 32 64 128 256 512 16 32 64 128 256 8 16 32 64 128 0.06 0.25 1 4 16
pg ml–1

b IL-6 IL-8 TNF-α IL-1β

0.4 P = 0.0017, HR = 2.0
0.4 P = 0.4883, HR = 1.1
0.4 P = 0.0005, HR = 2.2
0.4 P = 0.5831, HR = 0.9
Cumulative death

0.3 0.3 0.3 0.3

High
incidence

0.2 0.2 0.2 0.2

Low
0.1 0.1 0.1 0.1

0.0 0.0 0.0 0.0

0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
Days from ELLA cytokine test to last follow-up

c
Cytokine levels and O2 saturation Cytokine levels and severity
>95% (normal) <90% (very low) Moderate Severe COVID-19 with EOD
100 100 HR = 1.9, P = 0.0090
100 100 HR = 1.7, P = 0.0963
Survival probability

IL-6 75 75 75 75
High
50 50 50 50
Low HR = 3.4, P = 0.0021 HR = 1.8, P = 0.0173
25 25 25 25
IL-6 lo, O2SAT_MIN >95 IL-6 lo, O2SAT_MIN <90 IL-6 lo, moderate COVID-19 IL-6 lo, severe with EOD
IL-6 hi, O2SAT_MIN >95 IL-6 hi, O2SAT_MIN <90 IL-6 hi, moderate COVID-19 IL-6 hi, severe with EOD
0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
100 100 100 100 HR = 1.7, P = 0.0240
HR = 2.0, P = 0.0007
Survival probability

TNF-α 75 75 75 75
High
50 50 50 50
Low
25 HR = 4.2, P < 0.0001 25 25 HR = 2.4, P = 0.003 25
TNF-α lo, O2SAT_MIN >95 TNF-α lo, O2SAT_MIN <90 TNF-α lo, moderate COVID-19 TNF-α lo, severe with EOD
TNF-α hi, O2SAT_MIN >95 TNF-α hi, O2SAT_MIN <90 TNF-α hi, moderate COVID-19 TNF-α hi, severe with EOD
0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Days from ELLA cytokine test to last follow-up

Fig. 4 | Cytokine levels correlate with severity and independently predict survival. Correlation of cytokine levels with established inflammatory and
severity measurements. a, Correlation of each cytokine with each metric (n = 1,106 for fever, n = 1,112 for O2 saturation, n = 1,023 for CRP, n = 926 for
D-dimer, n = 1,017 for ferritin, n = 1,038 for platelets and n = 1,023 for disease severity score), using the same univariate and multivariate analyses as in
the Fig. 2 legend. Error bar indicates the median ± 95% CI. b, Competing risk analysis (n = 671) showing survival differences by IL-6 and TNF-α levels,
after adjusting the following variables: IL-6, IL-8, TNF-α, IL-1β, age, sex, race/ethnicity, smoking status, asthma, atrial fibrillation, cancer, CHF, CKD, COPD,
diabetes, hypertension, sleep apnea, severity, systolic blood pressure max, O2 saturation min, D-dimer, albumin, calcium, chloride and platelet count.
c, Kaplan–Meier univariate analyses of survival by IL-6 and TNF-α levels in patients with normal (n = 257), low (n = 258) or very low (n = 287) O2
saturation, or in patients with moderate (n = 588) versus severe COVID-19 with end organ damage (n = 136), as measured at the first available test.
EOD, end organ damage.

Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine 1641

Articles Nature Medicine

a Hydroxy- b
Tocilizumab Corticosteroids Remdesivir Acetaminophen chloroquine Anticoagulants Dexamethasone
4,096 4,096 4,096 4,096 4,096 4,096 4,096 Dexamethasone No
Tocilizumab No Steroids No Remdesivir No Acetomiphen No Hydroxychloroquine No Anticoagulants No
Tocilizumab Yes Steroids Yes Remdesivir Yes Acetomiphen Yes Hydroxychloroquine Yes Anticoagulants Yes 1,024 Dexamethasone Yes
1,024 1,024 1,024 1,024 1,024 1,024 256

–1

–1

–1

–1

–1

–1
–1

64

pg ml

pg ml

pg ml

pg ml

pg ml

pg ml
pg ml

IL-6 256 256 256 256 256 256 IL-6

16
64 64 64 64 64 64 4
16 16 16 1
16 16 16
0 5 10 15 20
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Time from tocilizumab Time from first encounter Time from remdesivir Time from first encounter Time from hydroxychloroquine Time from first encounter Time from first encounter
or first encounter or first encounter or first encounter
Methylprednisolone
1,024 Tocilizumab No
1,024 Steroids No
1,024 Remdesivir No
1,024 Acetomiphen No 1,024 Hydroxychloroquine No
1,024 Anticoagulants No 4,096 Methylprednisolone No
Tocilizumab Yes Steroids Yes Remdesivir Yes Acetomiphen Yes Hydroxychloroquine Yes Anticoagulants Yes
1,024 Methylprednisolone Yes
256 256 256 256 256 256 256
–1

–1

–1

–1

–1

–1

–1
IL-8
pg ml

pg ml

pg ml

pg ml

pg ml

pg ml

pg ml
IL-6 64
64 64 64 64 64 64 16
4
16 16 16 16 16 16 1
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Time from tocilizumab Time from first encounter Time from remdesivir Time from first encounter Time from hydroxychloroquine Time from first encounter Time from first encounter
or first encounter or first encounter or first encounter
Prednisone
256 256 256 256 256 256
Toci N Steroids No Remdesivir No Acetomiphen No Hydroxychloroquine No Anticoagulants No 4,096 Prednisone No
Toci Y Steroids Yes Remdesivir Yes Acetomiphen Yes Hydroxychloroquine Yes Anticoagulants Yes
1,024 Prednisone Yes
64 64 64 64 64 64 256
–1

–1

–1

–1

–1

–1

–1
pg ml

pg ml

pg ml

pg ml

pg ml

pg ml

pg ml
TNF-α 64
16 16 16 16 16 16 IL-6 16
4
4 4 4 4 4 4 1
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Time from tocilizumab Time from first encounter Time from remdesivir Time from first encounter Time from hydroxychloroquine Time from first encounter Time from first encounter
or first encounter or first encounter or first encounter

Fig. 5 | Treatment effect on cytokine levels. a, Effect of treatments on IL-6 (top row), IL-8 (middle row) and TNF-α (bottom row). Lines (in red: with
indicated treatment; in blue: without indicated treatment) represent the best fit curve by smoothed spline of the longitudinal and unique time point
distribution of each cytokine level based on time from either first encounter or treatment start. Of 1,670 samples representing various time points of 1,315
patients with available information, the number of those from patients who received tocilizumab, corticosteroids (any of prednisone, methylprednisolone
or dexamethasone), remdesivir, acetaminophen, hydroxychloroquine and/or anticoagulants (apixaban, enoxaparin, heparin or rivaroxaban) was 73, 305,
76, 620, 1,333, and 1,113, respectively. b, Effect of different corticosteroids on IL-6.

showed that these patients started with elevated IL-6 levels and then organ damage. Furthermore, elevated TNF-α, known to contribute
had a transient increase in serum IL-6, which has previously been to organ damage, was also a strong predictor of poor outcome even
explained by disrupted clearance after drug saturation of the IL-6 after adjusting for other risk factors such as age, sex, hypoxia, disease
receptor21. This transient elevation was observed only for IL-6, not severity scoring based on clinical assessment and IL-6. Our cyto-
IL-8, whereas TNF-α appeared to gradually decrease after therapy. kine panel also included IL-8, which showed association with sur-
Patients treated with corticosteroids and remdesivir showed, respec- vival time, even though it was eclipsed by other severity factors after
tively, a rapid and gradual reduction in IL-6 over time compared to multivariate adjustments, and IL-1β, which was poorly detected
patients who did not receive these drugs, but we observed no effect and, as a result, had only marginal predictive value. Although clas-
on TNF-α. Hydroxychloroquine, acetaminophen or anti-coagulants sic markers used routinely to determine inflammation and severity
did not clearly appear to alter cytokine levels. Of corticosteroids, were still useful to stratify patients on their own, when combined in
dexamethasone had the highest reduction effect on IL-6 (Fig. 5b), multivariate analyses, many were no longer significant, likely due to
potentially supporting findings from the recent RECOVERY trial collinearity, whereas IL-6 and TNF-α remained independently pre-
showing clinical benefit from this drug in hospitalized patients with dictive of outcome. Both overall survival and competing risk mod-
severe disease22. els used here consistently showed the significant prognostic value
of TNF-α and IL-6 when all tested cytokines were in the model,
Discussion along with demographics, comorbidities and other clinical and
We aimed to understand the role of inflammatory cytokines on laboratory measurements, highlighting the robustness of our find-
COVID-19 disease course and outcome. We established a rapid ings. Notably, the COVID-19-related cytokine response was quite
multiplex cytokine test to measure IL-6, TNF-α and IL-1β, as known distinct from the traditional cytokine storm associated with sepsis
markers of inflammation and organ damage, along with CXCL8/ and CAR T cells, with sustained elevated cytokine levels over days
IL-8 because of its potent role in the recruitment and activation of and weeks, and relative absence of coordination between cytokines.
neutrophils, commonly elevated in patients with COVID-19 (ref. 23). This raises the possibility of mitigation strategies with anti-cytokine
Notably, drugs blocking these cytokines are either FDA approved treatments, although which one(s) and the window of opportunity
or in clinical trials. Studying over 1,400 hospitalized patients in a for their use remain to be established. Guiding such therapies based
month, we established that COVID-19 is associated with high levels on mechanistic association with cytokine levels could provide a
of all four cytokines at presentation. Importantly, our observations rational approach.
indicate that cytokine patterns are predictive of COVID-19 survival Trials to block IL-6 signaling with already FDA-approved
and mortality, independently of demographics and comorbidities, drugs have been launched across the world, and some clinical
but also of standard clinical biomarkers of disease severity, includ- benefits have been seen in a subset of patients in small, single-
ing laboratory and clinical factors. A model based on these observa- center, observational studies15,24. In contrast, interim analysis of
tions was confirmed in a validation cohort of another 231 patents. randomized trials with the anti-IL-6 receptor monoclonal anti-
We found that IL-6 was one of the most robust prognostic markers body sarilumab versus placebo identified potential benefit only
of survival, eclipsing or outperforming CRP, D-dimer and ferritin in patients with severe but not moderate disease (https://investor.
after adjusting for the demographic features and comorbidities. It regeneron.com/news-releases/news-release-details/
remained independently associated with severity and predictive of regeneron-and-sanofi-provide-update-us-phase-23-adaptive).
outcome when including information about ventilation and end There are no available data correlating levels of IL-6 and response

1642 Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine

Nature Medicine Articles
to treatment, and none of the current studies has used cytokine 4. Merad, M. & Martin, J. C. Pathological inflammation in patients with
profiling as part of its inclusion criteria. It is possible that patients COVID-19: a key role for monocytes and macrophages. Nat. Rev. Immunol.
20, 355–362 (2020).
with moderate disease and high IL-6 levels will benefit the most 5. Vabret, N. et al. Advancing scientific knowledge in times of pandemics.
from cytokine blockade. Additionally, IL-6 reduction observed in Nat. Rev. Immunol. 20, 338 (2020).
patients treated with dexamethasone might be a mechanism under- 6. Mehta, P. et al. COVID-19: consider cytokine storm syndromes and
lying the efficacy of this treatment25. There is also a need to evaluate immunosuppression. Lancet 395, 1033–1034 (2020).
7. Ruan, Q., Yang, K., Wang, W., Jiang, L. & Song, J. Clinical predictors of
the effect of anti-TNF-α therapy on its own in COVID-19. Because
mortality due to COVID-19 based on an analysis of data of 150 patients from
IL-6 and TNF-α appear to be independent variables, studies with a Wuhan, China. Intensive Care Med. 46, 846–848 (2020).
combination regimen blocking both cytokines would be important 8. Wu, C. et al. Risk factors associated with acute respiratory distress syndrome
to consider for added clinical efficacy. and death in patients with coronavirus disease 2019 pneumonia in Wuhan,
Early cytokine measurements are reliable predictors of China. JAMA Intern. Med. https://doi.org/10.1001/jamainternmed.2020.0994
(2020).
outcome and, therefore, raise the critical importance of using 9. Herold, T. et al. Elevated levels of IL-6 and CRP predict the need for
serum cytokine levels for treatment decisions. The predictive mechanical ventilation in COVID-19. J. Allergy Clin. Immunol. 146,
value of these cytokines might help inform therapeutic interven- 128–136 (2020).
tions to determine which individuals are likely to develop respira- 10. Yang, Y. et al. Exuberant elevation of IP-10, MCP-3 and IL-1ra during
tory failure, end organ damage and death and to select optimal trial SARS-CoV-2 infection is associated with disease severity and fatal outcome.
Preprint at https://doi.org/10.1101/2020.03.02.20029975 (2020).
designs to disrupt the underlying inflammatory milieu. A predic- 11. Zhang X. et al. Viral and host factors related to the clinical outcome of
tion model built on cytokine levels early in disease might serve to COVID-19. Nature 583, 437–440 (2020).
inform healthcare allocation and prioritization of individuals at 12. Fehr, A. R., Channappanavar, R. & Perlman, S. Middle East respiratory
highest risk. syndrome: emergence of a pathogenic human coronavirus. Annu. Rev. Med.
Although confirmed in our validation cohort, the predictive 68, 387–399 (2017).
13. Moore, J. B. & June, C. H. Cytokine release syndrome in severe COVID-19.
value of IL-6 and TNF-α should also be assessed in a prospective Science 368, 473–474 (2020).
manner, where more control over data collection can be applied. 14. Davila, M. L. et al. Efficacy and toxicity management of 19-28z CAR T cell
Although the focus of this study was on only four cytokines chosen therapy in B cell acute lymphoblastic leukemia. Sci. Transl. Med. 6,
for their known inflammatory or pathogenic properties, additional 224ra225 (2014).
soluble analytes will likely be useful to consider to refine the sur- 15. Xu, X. et al. Effective treatment of severe COVID-19 patients with
tocilizumab. Proc. Natl Acad. Sci. USA 117, 10970–10975 (2020).
vival predictive model. Our current efforts are to build such a pre- 16. Feldmann, M. et al. Trials of anti-tumour necrosis factor therapy for
dictive model that will make use of a prospectively collected cohort COVID-19 are urgently needed. Lancet 395, 1407–1409 (2020).
where we will leverage high-dimensional assays such as Olink prox- 17. Dinarello, C. A., Simon, A. & van der Meer, J. W. Treating inflammation by
imity extension assay and the SomaLogic aptamer platform26, which blocking interleukin-1 in a broad spectrum of diseases. Nat. Rev. Drug Discov.
can measure hundreds to thousands of soluble analytes from serum 11, 633–652 (2012).
18. Baggiolini, M., Walz, A. & Kunkel, S. L. Neutrophil-activating peptide-1/
or plasma. The most informative dimensions from these assays interleukin 8, a novel cytokine that activates neutrophils. J. Clin. Invest. 84,
could then be carried back into the rapid 4-8 plex ELLA cytokine 1045–1049 (1989).
detection system for clinical decision-making, in addition to IL-6 19. Jatiani, S. S. et al. Myeloma CAR-T CRS management with IL-1R antagonist
and TNF-α. We think that these practices will bring cytokine mea- anakinra. Clin. Lymphoma Myeloma Leuk. https://doi.org/10.1016/j.
clml.2020.04.020 (2020).
surements to standard of care in prognosticating and monitoring
20. Madduri, D., Dhodapkar, M. V., Lonial, S., Jagannath, S. & Cho, H. J. SOHO
patients with COVID-19. State of the Art updates and next questions: T-cell-directed immune therapies
for multiple myeloma: chimeric antigen receptor-modified T cells and bispecific
Online content T-cell-engaging agents. Clin. Lymphoma Myeloma Leuk. 19, 537–544 (2019).
Any methods, additional references, Nature Research report- 21. Nishimoto, N. et al. Mechanisms and pathologic significances in increase in
serum interleukin-6 (IL-6) and soluble IL-6 receptor after administration of
ing summaries, source data, extended data, supplementary infor- an anti-IL-6 receptor antibody, tocilizumab, in patients with rheumatoid
mation, acknowledgements, peer review information; details of arthritis and Castleman disease. Blood 112, 3959–3964 (2008).
author contributions and competing interests; and statements of 22. The RECOVERY Collaborative Group. Dexamethasone in hospitalized
data and code availability are available at https://doi.org/10.1038/ patients with COVID-19 — preliminary report. N. Engl. J. Med. https://doi.
s41591-020-1051-9. org/10.1056/NEJMoa2021436 (2020).
23. Liu, J. et al. Longitudinal characteristics of lymphocyte responses and
cytokine profiles in the peripheral blood of SARS-CoV-2 infected patients.
Received: 28 May 2020; Accepted: 4 August 2020; EBioMedicine 55, 102763 (2020).
Published online: 24 August 2020 24. Gritti, G. et al. Use of siltuximab in patients with COVID-19 pneumonia
requiring ventilatory support. Preprint at https://doi.
References org/10.1101/2020.04.01.20048561 (2020).
1. Richardson, S. et al. Presenting characteristics, comorbidities, and outcomes 25. Ledford, H. Coronavirus breakthrough: dexamethasone is first drug shown to
among 5700 patients hospitalized with COVID-19 in the New York City area. save lives. Nature 582, 469 (2020).
JAMA 323, 2052–2059 (2020). 26. Rohloff, J. C. et al. Nucleic acid ligands with protein-like side chains:
2. Paranjpe, I. et al. Clinical characteristics of hospitalized COVID-19 patients modified aptamers and their use as diagnostic and therapeutic agents.
in New York City. Preprint at https://doi.org/10.1101/2020.04.19.20062117 Mol. Ther. Nucleic Acids 3, e201 (2014).
(2020).
3. Wang, B. et al. A tertiary center experience of multiple myeloma patients Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
with COVID-19: lessons learned and the path forward. J. Hematol. Oncol. 13, published maps and institutional affiliations.
94 (2020). © The Author(s), under exclusive licence to Springer Nature America, Inc. 2020

Nature Medicine | VOL 26 | October 2020 | 1636–1643 | www.nature.com/naturemedicine 1643

Articles Nature Medicine

Methods practice, which defined categories as follows: 1) mild/moderate COVID-19,

ELLA cytokine test. The ELLA platform is a rapid cytokine detection system based on based on normal/abnormal (<94%) O2 saturation, respectively, or pneumonia
four parallel singleplex microfluidics ELISA assays run in triplicate within cartridges on imaging; 2) severe COVID-19, based on use of high-flow nasal cannula
following the manufacturer’s instructions. We first validated IL-6, IL-8 and TNF-α (HFNC), non-rebreather mask (NRB), bilevel positive airway pressure (BIPAP) or
detection by ELLA at the Mount Sinai Human Immune Monitoring Center using mechanical ventilation and no vasopressor use, and based on CrCl greater than 30
plasma from multiple patients with myeloma who were undergoing immunotherapies and alanine aminotransferase (ALT) less than 5× the upper limit of normal; and 3)
such as CAR T cells and bispecific antibodies, known to elicit cytokine release storm. severe COVID-19 with end organ damage, based on use of HFNC, NRB, BIPAP or
Analytical validation (Extended Data Fig. 1a–c) was performed using both reference mechanical ventilation with use of vasopressors, or based on CrCl less than 30, new
cytokine controls and biological replicates across different lots of cartridges. The renal replacement therapy (hemodialysis/continuous veno-venous hemofiltration)
reproducibility was greater than 95%, with an intra-assay CV of 0.8%, an inter-assay or ALT more than 5× the upper limit of normal. Clinical notes and imaging reports
CV of 0.4–0.8% for analytes in the high detection range (>250 pg ml−1) and a CV were reviewed in an effort to establish the patients’ COVID-19 disease severity
of 2.6–4.2% for analytes in the lowest detection range (5–50 pg ml−1). Serum and over time. Using a bag-of-words approach to vectorize both clinical notes and
plasma appeared to be equivalent for detection of these cytokines. In March 2020, as image reports, vectors were derived from chest X-ray imaging reports, to reflect
the number of COVID-19 cases was increasing in New York City, we transferred the the presence of viral pneumonia and worsening respiratory symptoms. Intubation
ELLA methodology to the Mount Sinai Hospital Center for Clinical Laboratories, status and O2 therapy modality were obtaining by examining the patient clinical
which allowed the ELLA cytokine test to be coded into our electronic health record notes. The use of endotracheal tube, BIPAP, continuous positive air pressure,
ordering system as part of a COVID-19 diagnostic panel. HFNC, mechanical ventilator and/or supplemental O2 greater than FiO2 70% was
associated with severe COVID-19. End organ damage was defined by an ALT level
Patient information and data source. This research was reviewed and approved greater than 5× the upper limit of normal, CrCl less than 30, use of vasopressors
by the Human Research Protection Program at the Icahn School of Medicine at and/or new renal replacement therapy.
Mount Sinai (ISMMS). The Program for the Protection of Human Subjects is a key Alternatively, because of the prevalence in the literature, we also calculated the
component of ISMMS’ efforts to ensure human subject protections. It supports SOFA score (https://www.mdcalc.com/sequential-organ-failure-assessment-
our researchers in assuring the ethical conduct of research and compliance with sofa-score) for each visit, even if it normally best applied to those patients
federal, state and institutional regulations and provides a professional office staff to within the intensive care unit. SOFA score was significantly correlated with our
assist investigators, participants and five IRBs. A waiver of informed consent was hospital-based severity score (n = 1,450 pairs, Spearman’s r = 0.43, P < 0.0001).
obtained to query the patient electronic health records. Samples for the RT–PCR
SARS-CoV-2 lab test were collected via nasopharyngeal or oropharyngeal swab Statistical analysis. Patient characteristics were summarized using the standard
at one of 53 different Mount Sinai locations, representing outpatient, urgent care, descriptive statistics: median/IQR for continuous variables and count/percent for
emergency and inpatient facilities. Blood specimens for ELLA were collected via categorical variables. Distributions of the cytokine values were assessed and log2
venipuncture within the Mount Sinai Health System. All specimens and imaging transformed to render the parametric statistical analyses. The cytokines were then
were collected as part of standard of care. categorized at the level of 70 pg ml−1, 50 pg ml−1, 35 pg ml−1, 0.5 pg ml−1, 100 mg
Between March 21 and April 28, 2020, 1,484 patients hospitalized with L−1, 1,000 µg L−1 and 1 mg L−1 for IL-6, IL-8, TNF-α, IL-1β, CRP, ferritin and
suspicion of COVID-19 were tested for SARS-CoV-2 viral infection status by PCR D-dimer, respectively. These stringent cutoffs for cytokines were decided based on
and for the ELLA cytokine panel, and routine laboratory measurements and blood empiric testing of various cutoffs within COVID-19 samples, and choosing those
counts were obtained as part of standard medical care. For validation purposes, rounded above the median of COVID-19 distribution, except for TNF-α where
we also obtained data from an independent cohort of clinically annotated we chose a cutoff based on the upper 99th percentile value of controls because
SARS-CoV-2 PCR-positive hospitalized patients at Mount Sinai in whom cytokine of greater overlap in detection range, whereas those for elevated inflammatory
testing was performed between April 22 and June 16, 2020 (n = 231; median markers were based on 2–3× the upper limit of normal detection. The univariate
follow-up, 11.6 d, up to 53 d). Patients were identified by querying the pathology analyses assessed the association of the cytokines and laboratory tests with patient
department electronic database for individuals with both SARS-CoV-2 PCR-based characteristics using the Mann–Whitney U test, the Kruskal–Wallis test and
testing and ELLA cytokine panel. Cytokine data were obtained from pathology Spearman’s rank correlation test as appropriate. Additionally, Deming regressions
department electronic databases, and clinical and demographic data were and Spearman’s correlation coefficients were calculated for correlations between
supplemented with information from the Mount Sinai Data Warehouse. the cytokines and laboratory tests. We used multivariable linear regression models
A list of medical record numbers for patients who had both a SARS-CoV-2 to test the association of the cytokine values with patient demographics and
PCR result and an ELLA cytokine panel result in the pathology department comorbidities28. Kaplan–Meier plots along with log-rank tests were conducted
electronic database was provided to the Mount Sinai Data Warehouse. to assess the differences in survival probabilities between the high and low levels
Subsequently, demographic and clinical data were extracted from the Epic of each cytokine across the follow-up timeframe, which was calculated from the
electronic health record for the identified patients using the Epic Hyperspace date of cytokine testing to date of death, discharge or end of follow-up period
(August 2019), Epic Clarity (February 2020) and Epic Caboodle (February 2020) as appropriate29. The Cox proportional hazards model was used to estimate the
databases via connecting to Oracle (18c Enterprise Edition Release 18.0.0.0.0) and hazard of death adjusting for the covariates (for example, patient demographics,
SQL server (Microsoft SQL Server 2016 (SP2-CU11) (KB4527378) - 13.0.5598.27 comorbidities and laboratory test results), which were determined by the backward
(X64)) databases, respectively. Additional data elements included lab results, vital elimination method30,31. We assessed the survival model, censoring patients
signs, O2 therapy, radiology reports for chest imaging, diagnostic outcomes and discharged alive and in hospitals. The competing risk model, in which death was
medications. Data were merged from the various data sources using R version the event of interest, live discharge was the competing event and inpatients were
3.6.1. Large tables were read-in and written using the R packages tidyverse (v. censored32,33, was also fitted as a sensitivity analysis. Point estimates (HRs), along
1.3.0), reshape2 (v. 1.4.4) and readxl (v. 1.3.1) with the corresponding 95% CIs, predicted survival probabilities and cumulative
Clinical follow-up data were collected up to May 7, 2020, for the main cohort incidence curves, were provided. The analyses were performed using two-sided
and to June 23, 2020, for the validation cohort. Two investigators (D.M.D.V. and tests and the GraphPad Prism 8.4.2., SAS 9.4 and R 3.6.3 programs.
S.G.) independently compiled all clinical and laboratory information from these
various sources and compared them with near total matches. Differences were Validation approaches. We performed internal validation using two methods: 1)
adjudicated based on individual patient chart review and were explained by either ten-fold CV and 2) bootstrap validation (10,000 resamples). The discriminative
missing or updated information from the data warehouse. ability of the Cox proportional hazards regression model was assessed using
Harrell’s concordance index, denoting the probability that a randomly selected
Variables. Our data set included three broad classes of variables: 1) demographic patient with a higher survival time has a higher probability of survival predicted
variables (age, sex, race, ethnicity and smoking status); 2) clinical variables for compared to a randomly selected patient with a lower survival time. The goal is
each day of hospital encounter (BMI, heart rate, temperature, respiratory rate, to test the model’s ability to predict new data, to flag problems like overfitting
O2 saturation, systolic blood pressure, diastolic blood pressure, admission status, and selection bias and to give an insight on how the model will generalize to an
discharge status and deaths; and 3) comorbid conditions (CKD, asthma, COPD, independent external data set.
hypertension, obesity, diabetes, HIV, sleep apnea and cancer). All three categories In addition, we used an independent validation cohort (n = 231) and
were obtained from the patients’ electronic medical records, with comorbid applied the parameter estimates from the primary model (using demographic +
conditions defined as an active International Classification of Diseases (ICD)-10 comorbidity data as confounders) and the secondary model (using demographic
code and vital signs recorded for each patient’s given encounter. Although ICD-10 + comorbidity + lab data as confounders) to compute the prognostic index
codes represent an international system of categorical variables that are consistent (PI), the linear predictor derived from the model fitted to the derivation cohort.
between both practitioners and healthcare systems, we acknowledge that the Performance of this external validation model was captured by the models’
capture of these data from observational or retrospective cohorts might be less discrimination and calibration capabilities34. Discrimination was measured
reliable than prospective data collection focusing on specific data elements. by the AUCs over the follow-up time, which were estimated by the statistic c
(Harrell’s concordance index). Point estimates of 95% CIs and integrated AUCs
Determining COVID-19 disease severity. A severity scale for COVID-19 was were computed. Calibration was assessed by plotting Kaplan–Meier curves using
devised by pulmonologists at Mount Sinai based on literature27 and clinical the actual survival probabilities in the validation cohort and by comparing them

Nature Medicine | www.nature.com/naturemedicine

Nature Medicine Articles
with the corresponding predicted survival probabilities. Closeness of these two instruments and assay kits for ELLA testing in a Clinical Laboratory Improvement
curves is a sign of good calibration. The distribution of the PIs in the original and Amendments environment in the timeliest possible way during the health crisis; and
validation cohorts were presented as histograms and summarized in Extended R. Hyland from Gilead for giving permission to use the remdesivir-related data. S.G.,
Data Fig. 4. The similar spread of these distribution provides evidence toward the D.M.D.V., S.K.-S., Ma.M., H.-H.H., P.K., A.R. and Mi.M, were supported by National
appropriateness of the validation cohort. Cancer Institute U24 grant CA224319. S.G. is additionally supported by grants U01
DK124165 and P01 CA190174. T.H.S. was supported by National Institutes of Health
Reporting Summary. Further information on research design is available in the K08AI120806. A.R. was supported by U19 AI118610 and U24 AI1186441. Mi.M. was
Nature Research Reporting Summary linked to this article. supported by the Fast Grant fund. The Human Immune Monitoring Center and the
Institute for Healthcare Delivery Science received support from Cancer Center P30 grant
CA196521. S.P. was supported by grant R01 CA244899.
Data availability
The data supporting this publication have been made available at ImmPort
(https://www.immport.org) under study accession SDY1662. The data set has Author contributions
been de-identified in compliance with the Health Insurance Portability and D.M.D.V. and S.K.-S. contributed equally. H.-H.H. and N.D.B. contributed equally.
Accountability Act. ImmPort is a data sharing and data analysis portal for the D.M.D.V., S.K.-S., A.R., J.A.A., A.F.-B., D.R.M., J.J., D.R., C.C.-C., S.P., Mi.M. and S.G.
immunology research community funded by the National Institute of Allergy and contributed to study concept and design. D.M.D.V., S.K.-S., H.-H.H., N.D.B., S.N., B.W.,
Infectious Diseases and the Division of Allergy, Immunology and Transplantation. Y.L., T.H.S., D.M., A.S., T.U.M., O.V.O., A.R., P.K., J.A.A., E.S., S.J., Ma.M., A.W.C.,
For further details, refer to the ImmPort user agreement (https://www.immport. A.F.-B., D.R.M, J.J., D.R., K.S., M.F., C.C.-C., S.P., Mi.M. and S.G. contributed to literature
org/agreement). search, writing the manuscript and data interpretation. D.M.D.V., S.K.-S., H.-H.H.,
N.D.B., S.N., B.W., Y.L., T.H.S., D.M., A.S., T.U.M., X.H., M.P., K.T., O.V.O., A.R., P.K.,
Ma.M., A.W.C., A.F.-B., D.R.M., J.J., D.R., K.S., C.C.-C., S.P., Mi.M. and S.G. participated
Code availability
in data collection and data analysis. D.M.D.V., S.K.-S., H.-H.H., N.D.B., Ma.M., K.S. and
Scripts used to query the Clarity and Caboodle databases, as well as the statistical
S.G. made figures and tables.
analysis, have been uploaded to a GitHub repository: https://github.com/delvad03/
COVID19ELLA.
Competing interests
S.G. reports consultancy and/or advisory roles for Merck, Neon Therapeutics and
References OncoMed and research funding from Bristol-Myers Squibb, Genentech, Immune Design,
27. Cascella, M., Rajnik, M., Cuomo, A., Dulebohn, S. C. & Di Napoli, R. Agenus, Janssen R&D, Pfizer, Takeda and Regeneron. S.P. reports consulting fees from
Features, evaluation and treatment coronavirus (COVID-19). In StatPearls Foundation Medicine and research funding from Celgene and Karyopharm. D.M.
(StatPearls Publishing, 2020). reports consultancy and/or advisory roles from Janssen, Celgene, Bristol-Myers Squibb,
28. Harrell, F. Regression Modeling Strategies (Springer, 2001). Takeda, Legend, GlaxoSmithKline, Kinevant and Foundation Medicine. B.W. reports a
29. Rich, J. T. et al. A practical guide to understanding Kaplan–Meier curves. consulting role at Sanofi Genzyme. J.A.A. reports grants and personal fees from Gilead,
Otolaryngol. Head. Neck Surg. 143, 331–336 (2010). grants and personal fees from Merck, grants and personal fees from Janssen, personal
30. Cox, D. Analysis of Survival Data (Chapman & Hall, 1984). fees from Theratech, personal fees from Medicure, grants from Regeneron and grants
31. Smith, T. & Smith, B. Survival analysis and the application of Cox’s and personal fees from ViiV, all outside of the submitted work. S.J. reports consulting
proportional hazards modeling using SAS. In SAS Conference Proceedings and/or advisory roles from Janssen, Celgene, Bristol-Myers Squibb, Takeda, Legend and
(SAS Global Forum, 2001). GlaxoSmithKline.
32. Satagopan, J. M. et al. A note on competing risks in survival data analysis.
Br. J. Cancer 91, 1229–1235 (2004).
33. Berger, M., Schmid, M., Welchowski, T., Schmitz-Valckenberg, S. & Additional information
Beyersmann, J. Subdistribution hazard models for competing risks in discrete Extended data is available for this paper at https://doi.org/10.1038/s41591-020-1051-9.
time. Biostatistics 21, 449–466 (2018).
Supplementary information is available for this paper at https://doi.org/10.1038/
34. Royston, P. & Altman, D. G. External validation of a Cox prognostic model:
s41591-020-1051-9.
principles and methods. BMC Med. Res. Methodol. 13, 33 (2013).
Correspondence and requests for materials should be addressed to S.G.
Acknowledgements Peer review information Saheli Sadanand was the primary editor on this article and
Authors thank N. Fernandez for help with the clustergrammer tool and members of managed its editorial process and peer review in collaboration with the rest of the
the Human Immune Monitoring Center for help with specimen handling. The authors editorial team.
wish to acknowledge R. Pande and M. Putnam at Bio-Techne for helping to provide Reprints and permissions information is available at www.nature.com/reprints.

Nature Medicine | www.nature.com/naturemedicine

Articles Nature Medicine

Extended Data Fig. 1 | Sensitivity, specificity, and reproducibility testing of the ELLA platform and cytokine levels and correlation observed in COVID-19
specimens. a, Spearman r correlation between IL-6 tested by two other platforms for soluble analyte detection, Olink (n = 18) and LabCorp (n = 142), using
plasma or serum specimens from CAR-T CRS or COVID-19. b, Interassay and intraassay coefficient of variation (CV) of replicates for two recombinant
controls used at high or low concentration in each assay for IL-6 and TNF-α using a first set of controls, and for the ELLA panel used in this study using
Randox recombinant antigens at three dilution levels over 28 dates tested. c, Reproducibility testing replicates of the same biological specimen from
CAR-T samples, with Spearman r indicated for 40 paired samples for IL-6 and 8 paired samples for TNF-α. d, Distribution of each cytokine in all COVID-19
samples tested as shown in Fig. 1. e, Correlation matrix of IL-1β, IL-6, IL-8, and TNF-β levels in COVID-19 plasma specimens (n = 1,949) and in E. multiple
myeloma specimens during immunotherapy-related CRS (n = 121). Scale indicates value of Spearman r correlation. Cytokines levels are less coordinated in
COVID-19 than in CAR-T CRS.

Nature Medicine | www.nature.com/naturemedicine

Nature Medicine Articles

Extended Data Fig. 2 | Flow chart to determine severity. HFNC: high flow nasal cannula; NRB: non-rebreather mask; BiPAP: bilevel positive airway
pressure; CPAP: continuous positive airway pressure; SpO2: oxygen saturation; CrCl: creatinine clearance. ALT: alanine aminotransferase; ULN: upper level
of normal; HD/CVVH: hemodialysis / continuous venovenous hemofiltration; EOD: end organ damage.

Nature Medicine | www.nature.com/naturemedicine

Articles Nature Medicine

Extended Data Fig. 3 | See next page for caption.

Nature Medicine | www.nature.com/naturemedicine

Nature Medicine Articles
Extended Data Fig. 3 | Correlations of clinical and laboratory measurements in individual patients. a, Unsupervised clustering of laboratory
measurements in a subset of 1,069 patients with sufficient available information. On the y axis are vitals and laboratory values after z-scoring, and on
the x axis are individual patients, using metrics measured from the time point corresponding to the first ELLA cytokine test. Grey bars on the side of
the plot indicate clusters of patients or analytes, where cytokines co-cluster with known severity metrics, such as LDH, CRP, ferritin, D-dimer, but also
high neutrophil, platelet and white blood counts. Annotations show patients who died in orange, and maximum severity score achieved in gray shades.
b, Similarity matrix of patients based on analytes and measurements, showing two major clusters, with enrichment in patients who died and had more
severe COVID-19 on the upper left. c, Similarity matrix of cytokines, lab measurements and vitals, showing IL-6, IL-8, and IL-1β co-clustering with known
inflammatory markers such as LDH, CRP, ferritin, and D-dimer, while TNF-α co-clusters with organ damage markers.

Nature Medicine | www.nature.com/naturemedicine

Articles Nature Medicine

Extended Data Fig. 4 | Validation of the models in an independent cohort. Performance of the model including demographics and comorbidities (a-c)
or additionally including laboratory metrics (d-f) in a validation cohort of 231 patients. a and d, Pointwise time-dependent area under the curve (AUC)
along with the 95% confidence interval were computed to assess the discrimination of the model from cytokine test to last follow-up, with values well
above 0.5. b and e, Calibration was assessed by plotting Kaplan-Meier curves using the actual survival probabilities in the validation cohort and by
comparing it with the corresponding predicted survival probabilities. Closeness of these two curves is a sign of good calibration. c and f, The distribution
of the prognostic index in the original cohort (red) and the validation cohort (green) shown as histogram displays a similar spread, providing evidence
towards the appropriateness of the validation cohort. Median (IQR) for primary model without labs: original cohort: 0.2401 (-0.3874, 0.8168), validation
cohort: 0.0697 (-0.4785, 0.7722). Median (IQR) for model including labs: original cohort: -0.3671 (-1.3246, 0.5021), validation cohort: -0.6642 (-1.5348,
0.3444).

Nature Medicine | www.nature.com/naturemedicine

 nature research | reporting summary
Corresponding author(s): Sacha Gnja�c, PhD
Last updated by author(s): 7/28/2020

Repor�ng Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in repor�ng. For further informa�on on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Sta�s�cs
For all sta�s�cal analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods sec�on.
n/a Confirmed
The exact sample size (n) for each experimental group/condi�on, given as a discrete number and unit of measurement
A statement on whether measurements were taken from dis�nct samples or whether the same sample was measured repeatedly
The sta�s�cal test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A descrip�on of all covariates tested

A descrip�on of any assump�ons or correc�ons, such as tests of normality and adjustment for mul�ple comparisons
A full descrip�on of the sta�s�cal parameters including central tendency (e.g. means) or other basic es�mates (e.g. regression coefficient)
AND varia�on (e.g. standard devia�on) or associated es�mates of uncertainty (e.g. confidence intervals)

For null hypothesis tes�ng, the test sta�s�c (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, informa�on on the choice of priors and Markov chain Monte Carlo se�ngs
For hierarchical and complex designs, iden�fica�on of the appropriate level for tests and full repor�ng of outcomes
Es�mates of effect sizes (e.g. Cohen's d, Pearson's r), indica�ng how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

So�ware and code

Policy informa�on about availability of computer code
Data collec�on Data was processed using R version 3.6.1 (Vienna Austria). Large tables were read-in and wri�en using R packages �dyverse (v 1.3.0),
reshape2 (v 1.4.4) , and readxl (v 1.3.1)
Clinical data was extracted from the Epic electronic health record (Verona, WI) using Epic Hyperspace (August 2019), Epic Clarity (February
2020) and Epic Caboodle (February 2020) databases via connec�ng to Oracle (18c Enterprise Edi�on Release 18.0.0.0.0) and SQL server
(Microso� SQL Server 2016 (SP2-CU11) (KB4527378) - 13.0.5598.27 (X64)) databases respec�vely.

Data analysis SAS 9.4 (PROC GENMOD, PROC LIFETEST and PROC PHREG), GraphPad Prism 8.4.2, and R 3.6.3. (Package: ggplot2)
For manuscripts u�lizing custom algorithms or so�ware that are central to the research but not yet described in published literature, so�ware must be made available to editors and
reviewers. We strongly encourage code deposi�on in a community repository (e.g. GitHub). See the Nature Research guidelines for submi�ng code & so�ware for further informa�on.

Data
Policy informa�on about availability of data
April 2020

All manuscripts must include a data availability statement. This statement should provide the following informa�on, where applicable:
- Accession codes, unique iden�fiers, or web links for publicly available datasets
- A list of figures that have associated raw data
- A descrip�on of any restric�ons on data availability

The data suppor�ng this publica�on is available at ImmPort (h�ps://www.immport.org) under study accession SDY1662.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size A two-sided log-rank test with an overall sample size of 674 patients (337 in each group) achieves 80% power at a 0.05 type I error to detect a
hazard ratio of 2.29 when the proportion of death in the low group is 20%. In this study, we had more than 1,200 patients so that we can have
enough power to detect the potential slight smaller effect sizes after adjusting for the covariates. Please refer to response to Reviewer 2
comments. We queried all available health records.

Data exclusions The patients with the cytokines measured after the date of last follow-up and or missing status (discharge, death, or hospitalization) were
excluded. 1,298 patients had cytokines measured on or before the date of last follow-up. Among them, 1,245 patients had the measurements
of all four cytokines. Exclusion criteria were pre-established. Our aim was to look at patients with both an indication of SARS-CoV-2 infection
and ELLA cytokine data availability.

Replication This observational study is of a hospitalized cohort during the COVID-19 pandemic in New York City. We validated the initial findings in a
second cohort. -- The ELLA platform has 3 built in replicates.

Randomization This is not a clinical trial. Established clinical laboratory test data was used.

Blinding This is not a clinical trial.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
Dual use research of concern

Human research participants

Policy information about studies involving human research participants
Population characteristics Please see Table 1. Patient characteristics for details of human subjects research participants.

Recruitment Subjects were not recruited into this study, therefore, self-selection bias does not apply. Data was collected as part of
normal clinical care. ELLA was ordered as a clinical test as part of standard clinical care at the MSHS during the SARS-CoV-2
epidemic. Clinical tests were ordered by the treated physician.

Ethics oversight This research was reviewed and approved by the Human Research Protection Program at the Icahn School of Medicine at
April 2020

Mount Sinai (ISMMS), a comprehensive system to ensure the protection of the rights and welfare of subjects in human
research. The Program for the Protection of Human Subjects (PPHS) is a key component of ISMMS’ efforts to ensure human
subject protections. It supports our researchers in assuring the ethical conduct of research and compliance with federal,
state, and institutional regulations, and provides a professional office staff to assist both investigators, participants, and five
Institutional Review Boards (IRBs).
Note that full information on the approval of the study protocol must also be provided in the manuscript.