In this single-center retrospective cohort study, we extracted electronic medical records of all patients admitted to Tel Aviv Sourasky Medical Center, Israel, who had blood cultures drawn, or tested positive for a viral infection, made available between July 2007 and May 2023.

The outcome was patients’ CRP levels across time, considered in mg/L throughout. The time of CRP measurement was defined relative to each patient’s time zero: the time of the first bacterial blood culture or viral diagnostic test collection, and was measured up to 24 h prior to culture or viral test obtainment, and up to 150 h following it. Patients whose initial CRP level was higher than 31.9 were excluded. This exclusion criterion, constituting three standard deviations above normal CRP levels [6], was devised to remove individuals already in a highly active inflammatory process, so that standard CRP dynamics could be captured.

Data were stratified into three cohorts for the analysis (Supplementary Table 1): (1) a subsample of patients who had negative bacterial blood cultures (35,975 patients, 103,333 CRP measurements) and had not tested positive for viral infection, but likely had some infection/inflammation leading to drawing of culture; (2) patients who tested positive for any virus (Supplementary Table 2), while testing negative for bacteremia (1805 patients, 4426 CRP measurements); (3) a subsample of patients who had positive bacterial growth in their blood cultures. Cultures with organisms likely to be contaminants were excluded (Supplementary Tables 3–4, respectively). Patients who had growth of non-contaminant bacteria in any of the cultures drawn were stratified to those who only had a Gram-negative infection (718 patients, 2658 CRP measurements) and those who only had a Gram-positive infection (378 patients, 1473 CRP measurements). All patients were further classified as to whether they received any antibiotic treatment up to 24 h before or after time zero vs. either not having received antibiotics or having received them outside this window. Additional covariates included in analyses were patients’ Charlson comorbidity index (CCI), sex, and age.

We employed generalized additive mixed-effects models (GAMMs), using the R package mgcv [7], as our primary method of analysis. GAMMs are an extension to generalized linear mixed-effects models, where the assumption of a linear relationship of a covariate and the outcome is replaced with a smooth estimation through splines. These models allowed for flexible estimation of the relationship between the time and CRP trajectory, while controlling for age, sex, and CCI, and including a random intercept to account for patients’ repeated measures. A GAMM was fitted to each of the cohorts. For the third cohort, bacterial Gram stain was included as an additional covariate. In addition, because the sample of patients who simultaneously tested negative for both bacterial and viral infection was very large, we randomly selected 1000 patients who received antibiotic treatment within the time window and 1000 who did not - numbers chosen to be relatively comparable in size to the virus cohort.

While the GAMMs produce inferential statistics for the smooth terms and the fixed effects coefficients, we were also interested in testing whether the peak CRP observed, as well as the area under the CRP curve (AUC-CRP), differed between strata. Statistical inference (confidence interval and p-value calculations) was done by bootstrap analysis, computing both the peak CRP and the area under the curve within the first 72 h of observation in each of 1000 bootstrap samples for each of the analyses performed. Statistical significance was determined at a p < 0.05 level. Further details are in Supplementary Tables 5–6.