Methods
Patients and inclusion
criteria
Bichat-Claude Bernard hospital is a large 850-bed university reference
center for high-risk pathogens in Northern Paris. All consecutive
patients presenting with an ILI and admitted to the hospital through the
ED were prospectively included from November 18, 2019 to March 30, 2020.
All were tested for respiratory viral infection using mPCR assay (see
below). The first COVID-19 case at our ED was identified on February 28,
and systematic on-site specific SARS-CoV-2 RT-PCR began on March 2,
2020.
An ILI was defined as the association of at least one systemic symptom
(fever with temperature higher than 38.5°, malaise, headache and
myalgia) with one respiratory symptom (cough, sore throat and dyspnea),
according to the eCDC ILI definition 9. Those
indications of PCR testing were identical during both periods in our ED
despite the constraints of the epidemics.
Study periods
In France, the influenza epidemic period extended from November 18, 2019
to March 2, 2020. During this period, SARS-CoV-2 was not spreading in
the community or among patients returning or travelling from foreign
countries, notably Northern Italy and China. All suspicions of
SARS-CoV-2 based on these exposures were thus admitted from other EDs or
hospitals in the dedicated ward of the Infectious Diseases department
and were not included in this study. Starting on March 3, all patients
with an ILI visiting the ED were systematically suspected of COVID-19
and tested for both human RVs and SARS-CoV-2.
To reflect these two periods, they were analyzed separately: (i) Period
1, hereafter called the respiratory virus (RV) period, with active
circulation of non-SARS-CoV-2 RVs, from November 18, 2019 to March 2,
2020, and (ii) Period 2, hereafter called the SARS-CoV-2 period with
active circulation of SARS-CoV-2, starting on March 3.
Virological testing
During the RV period, multiplex PCR was performed as point of care using
the QIAstat-Dx Respiratory Panel (QIAGEN, Hilden, Germany). This rapid
multiplex PCR assay allows for the detection of 22 viral and bacterial
respiratory targets, including influenza, parainfluenza,
rhinoviruses/enteroviruses, RSV, metapneumovirus, adenovirus,
coronaviruses, bocavirus, Mycoplasma pneumoniae ,Legionella pneumophila and Bordetella pertussis10. During the SARS-CoV-2 period, and depending on kit
availability, viral investigations were conducted either with the
QIAstat-Dx Respiratory SARS-CoV-2 Panel (Qiagen, Hilden, Germany),
allowing for the detection of the same respiratory pathogens plus
SARS-CoV-2 11, or with a combination of the RT-PCR
RealStar SARS-CoV-2 Kit RUO (Altona Diagnostics, Hamburg, Germany)12 and rapid multiplex PCR FilmArray RP2 (BioFire,
BioMérieux, Marcy-L’Etoile, France) 13, allowing for
the detection of the same viral respiratory pathogens except bocavirus.
Clinical and biological
data
Demographic and clinical data were prospectively collected, including
age, sex, respiratory symptoms (e.g. cough, dyspnea, expectoration,
chest pain and auscultatory abnormalities), time since the onset of
symptoms, comorbidities (e.g. diabetes, history of stroke, myocardial
infarction, chronic heart failure, chronic kidney failure, chronic
bronchitis and asthma) and clinical parameters at ED admission (e.g.
temperature, blood pressure, cardiac frequency, respiratory frequency,
oxygen saturation and Glasgow coma score). The following biological data
were also recorded: blood counts, C-Reactive Protein and NT-pro brain
natriuretic peptide (NT pro-BNP). Based on the ongoing acquisition of
COVID-19-related symptoms, diarrhea, anosmia and ageusia were added to
data recording after the beginning of the SARS-CoV-2 period.
Modelling strategy for assessing the effect of the
co-circulation of SARS-CoV-2 and other respiratory
viruses
Under the hypothesis of a co-circulation of SARS-CoV-2 and other RVs, we
modelled the odds of having a positive PCR for SARS-CoV-2 with respect
to other viruses, given that the PCR is positive. For the sake of
clarity, we denoted the sample by S , with the convention S= 1 for the SARS-CoV-2 period and S = 2 for the RV period. We considered
a mixture SARS-CoV-2 and RV period with parameter\(\ p=P(S=1)\).
According to Bayes formula, the probability of a random observation from
the mixture actually being drawn from Sample 1 given that the PCR is
positive is as follows:
\begin{equation}
P\left(S=1\right|PCR+)=\frac{P\left(PCR+\right|S=1)\ p}{P\left(PCR+\right|S=1)p+P\left(PCR+\right|S=2)(1-p)}\nonumber \\
\end{equation}
and the corresponding odds are
\begin{equation}
O\left(S=1\right|PCR+)=\frac{P\left(PCR+\right|S=1)\ p}{P\left(PCR+\right|S=2)(1-\ p)}\nonumber \\
\end{equation}
Therefore, if we consider a binary covariate X, the odds ratio for being
COVID+ in the strata X = 1 with respect to the strata X = 0 can be
written as follows:
\begin{equation}
\text{OR}_{S,X|PCR+}=\frac{\text{RR}_{PCR,X|1}}{\text{RR}_{PCR,X|2}}\text{OR}_{S,X}\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (1)\nonumber \\
\end{equation}
where \(\text{RR}_{\text{PCR},X|s}\)denotes the relative risk of having
a positive PCR in the strata X = 1 with respect to X = 0, given that the
observation belongs to sample \(s\). This equality, initially derived in
the case of a binary covariate X, can be straight forwardly extended to
the case of categorical or quantitative variables and covariate-adjusted
quantities.
Equation (1) has a clear interpretation in the sense that it identifies
two independent contributions to the
quantity\(\ \text{OR}_{S,X|\text{PCR}+}\). The first one is the ratio
between two relative risks and reflects the difference of association
between the covariate and the PCR diagnosis in both samples. If, for
example, the male sex is strongly associated with a higher risk of
having a positive PCR for SARS-CoV-2, but no association exists for
other RVs, then a positive PCR drawn from the mixture sample tends to
more likely be for SARS-CoV-2 if the patient is male. The second
contribution comes from the imbalance between both samples for covariate
X. If, for example, patients from the SARS-CoV-2 period are
significantly younger, then a positive PCR in a young patient is more
likely to be a COVID case.
Statistical analysis
The baseline characteristics within each group were described with
numbers and percentages for qualitative variables and median and
interquartile range (IQR) for quantitative variables.
We assessed factors associated with a positive PCR for SARS-CoV-2 in the
SARS-CoV-2 period and for other RVs in the RV period using univariate
logistic regression. For each sample and variable, we reported the Odd
ratio (OR), the 95% confidence interval using a normal approximation
and the p-value corresponding to the Wald statistic. We used a threshold
of 0.05 for statistical significance. We then performed a multivariate
model resulting from a stepwise selection procedure among variables with
a p-value below 0.05 in the univariate analysis and a proportion of
missing values below 20%.
Regarding the modelling of the SARS-CoV-2 risk in a period of
co-circulation with other RVs, we estimated relative risks on the
right-hand side of Equation (1) with a quasi-Poisson regression model to
take over-dispersion into account. Since these relative risks vary
according to the proportion of positive PCR in both samples, we built
different scenarios corresponding to different degrees of prevalence of
positive PCR for SARS-CoV-2 and other RVs. Relative risks for these
scenarios were estimated with a-priori weights on the observations. The
quantity \(\text{OR}_{S,X}\) and its standard error were estimated with
a logistic regression from the union of both samples. Finally, the
estimate for \(\text{OR}_{S,X|\text{PCR}+}\) was reported, and its 95%
confidence interval was derived with a normal approximation under the
hypothesis of independence between its components.
We also performed a multivariate model for the quantity\(\text{OR}_{S,X|\text{PCR}+}\) by using multivariate models for the
right-hand side of (1) and a stepwise selection procedure based on
covariates with p < 0.05 in the univariate analysis and less
than 20% of missing data. Moreover, to evaluate the discrimination of
the multivariate model, we reported the AUC and its 95% confidence
interval, derived from 2,000 bootstrap replicates. We finally proposed a
clinical score from the multivariate analysis by rounding estimates, and
we assessed its discrimination performances (AUC, sensibility and
specificity for a given threshold). All analyses were performed using R
v4.0.2.
The study has been approved by our local ethic committee.