Study design and setting
We conducted a historical cohort study of patients admitted through the ED in Tel Aviv Sourasky Medical Center (TASMC), Israel, between April 2011 and March 2019. We proposed investigating our assumption by looking at patients with viral and bacterial infections who presented with relatively low first CRP measurement (CRP1) and received antibiotics soon after. This cohort enabled us to evaluate the differentiation ability of CRPv when the CRP1 upon admission was not indicative of infection type, and yet, physicians decided on treatment with antibiotics. This clinical scenario reflects a setting where a differentiation biomarker could be highly significant and prevent inappropriate antibiotic treatment. Hospital records were reviewed manually in order to apply the exclusion criteria.
Patients
All the patients had a relatively low CRP1 ≤ 31.9 mg/L, upon admission and a positive lab test for either a viral or bacterial infection. This relatively low CRP concentration cut-off represents the mean CRP of apparently healthy individuals + 3 standard deviation. It was suggested by Wasserman et al., based on a relatively large cohort of individuals who were screened as part of a routine annual check-up [8].
Inclusion criteria
We included all patients admitted to the general ED in TASMC, Israel, who are adults aged ≥ 18. We could not include younger patients in this study because they are admitted to a separate pediatric ED in TASMC, and we did not have access with the current Helsinki approval to their medical records. Bacterial infections were identified by a positive blood culture for a single bacteria species that is likely to cause infection and not be a contaminant (defined as blood culture results of Diphtheroids spp, Coagulase-negative Staphylococcus, Streptococcus viridans group and a result of Gram-positive bacilli). Viral infections were identified by either a positive PCR for a virus or an immunoglobulin test indicative of a single viral species.
To link between the CRP measurement and the infectious diagnosis we only included patients whose positive blood culture was taken within 1 day or positive viral test within 5 days of the first CRP measurement. The time difference between CRP1 to the second CRP measurement (CRP2) was less than 24 h.
Exclusion criteria
Patients who were not treated with antibiotics between CRP1 and CRP2 or received antibiotics during the 24 h before admission.
Patients with co-infections, defined as having both a bacterial and viral positive test.
Patients with active malignancy or active inflammatory disease, for example, Systemic lupus erythematosus, Inflammatory bowel disease etc. Patients who were treated with anti-inflammatory medications or immunocompromised patients. Pregnant women and patients with missing medical records were excluded as well (Fig. 1).
Laboratory methods
Wide range (WR) CRP measurements were done by ADVIA 2400 (Siemens Healthcare Diagnostics Inc., Tarrytown, NY, USA). The ADVIA® Chemistry WR-CRP method measures CRP in the serum and plasma by a latex-enhanced immunoturbidimetric assay.
Computational methods
CRP velocity (CRPv) was defined as the difference between the second CRP measurement (CRP2) and the first measurement (CRP1), divided by the time difference between the tests (Δt CRP1 to CRP2, hours).
$$\mathrm{C}RPv=\frac{CRP2-CRP1}{\Delta\mathrm{t \,\,CRP}1\,\,\mathrm{to\,\, CRP}2}$$
Estimated CRP (eCRP) was defined as expected CRP in apparently healthy individuals based on age and sex. eCRP was calculated based on the data of Tel Aviv Medical Center Inflammation Survey. First, we divided the cohort by sex and then we calculated the mean CRP by age groups of 5 years (starting with individuals 20–25 years old to 75 years). There was a relatively small number of subjects in the male and female group above the age of 75, hence, we calculated their mean CRP concentration and considered it as the eCRP of all the patients above the age of 75 years. In addition, the healthy cohort did not include subjects younger than 20 years, so we clinically estimated the eCRP value of this age group (Our cohort had only one female patient younger than 20 and her eCRP value was 1.5 mg/L). The eCRP values of each age group are reported on Additional file 1: Table S1. The purpose of calculating eCRP is to estimate the CRP level of the patient in the healthy condition in the initiation of his pathological course. Estimated CRP velocity (eCRPv) was defined as the difference between CRP1 and eCRP divided by the time difference between the beginning of symptoms and CRP1 test (Δt onset of symptoms to CRP1, hours) was defined as eCRPv.
The exact timing from the onset of symptoms was estimated based on the medical record of each patient’s admission file.
$$\mathrm{eC}RPv=\frac{CRP1-\mathrm{eCRP} }{\Delta \mathrm{t \,\,between \,CRP}1\,\mathrm{and\, symptoms \,onset}}$$
The recorded antibiotic time was used to calculate the Δt Antibiotic; the time difference between the CRP1 measurement and antibiotic administration.
$$\Delta\mathrm{t \,Antibiotic}={{t}_{Antibiotic \, Administration}}-{{t}_{CRP1}}$$
For cases in which the administration time of antibiotic in ED was missing, we considered the time of admission to ward to be the antibiotic administration time.
Statistical analysis
Categorical variables were reported as numbers and percentages. The continuous variables were reported as means with standard deviations. To compare the distributions of different features on the bacterial and the viral groups we used the non-parametric test of Mann–Whitney (MW). In order to compare the sex categorical feature, we used the Pearson’s Chi-squared test (Chi2). To test the performance of the numerical parameters (e.g., eCRP, CRP1, CRP2, \(\mathrm{eC}RPv\), \(\mathrm{C}RPv\), etc.) as biomarkers for classification, we used receiver operator characteristic curve (ROC) analysis and calculated the area under the curve (AUC) and for each parameter. Standard non-parametric bootstrapping of 1000 samples was used to generate the 95% confidence intervals. The statistical analysis was performed using the Python programming language version 3.5.2 and the packages SciPy, NumPy and scikit-learn.