Longitudinal measurements of inflammatory mediators in patients at risk of sepsis in the Emergency Department

Abstract

Sepsis is a complex condition with time‑dependent evolution, and longitudinal biomarker dynamics may improve its characterisation. We hypothesised that biomarker kinetics are associated with sepsis, the intensity of organ dysfunction, and patient survival.

This single‑centre, prospective, observational study included adult patients presenting to the Emergency Department with suspected infection and a National Early Warning Score 2 (NEWS2) ≥3. Blood samples were collected at baseline, 4 and 24 h and pro-inflammatory and anti-inflammatory cytokines, chemokines, as well as other inflammatory mediators or acute-phase reactants, were analysed. Linear and nonlinear mixed-effects models assessed associations between biomarker trajectories over time, sepsis diagnosis, and organ dysfunction severity. Joint models evaluated the predictive value of biomarker kinetics during the first 24 h for in‑hospital mortality.

Of 214 screened patients, 173 were analysed and 137(79%) developed sepsis. Linear and nonlinear mixed models showed significant time‑dependent decreases in IL10 (β −0.016, 95%CI −0.028 to −0.004), IL1RN (β −0.014, 95%CI −0.024 to −0.004), and IL6 (β −1.021, 95%CI −1.609 to −0.433 for the first spline 0−4 h). Sepsis diagnosis was associated with higher IL1RN (β 0.378, 95%CI 0.153–0.603) and TNFRSF1A (β 0.40, 95%CI 0.21–0.58); a significant interaction between sepsis and time was observed only for IL6 (β −0.33, CI95% −0.653 to-0.007) for the second spline 4–24 h.

Increasing SOFA scores correlated with elevated IL10 (β 0.048), IL1RN (β 0.044), CCL2 (β 0.046), TNFRSF1A (β 0.050), and procalcitonin (PCT; β 2.15), all p < 0.01. The SOFA score–time interaction was significant only for IL6 (β 0.122, 95%CI 0.006 to 0.238 during the first spline 0–4 h followed by β −0.080, 95%CI −0.133 to −0.027 for the second spline 4–24 h). Joint survival models adjusted identified associations between longitudinal IL8 (HR 0.655, 95%CrI 0.582–0.728), TNFRSF1A (HR 0.505, 95%CrI 0.419–0.682), and PCT (HR 1.004, 95%CrI 1.001–1.008) concentrations and survival outcome. However, IL8 and TNFRSF1A showed an unexpected inverse association with survival, and the overall discriminatory performance was modest.

In conclusion, Sepsis status and organ dysfunction severity were associated with distinct inflammatory biomarker levels and kinetics. IL6 showed a unique time‑dependent relationship, whereas the dynamics of TNFRSF1A, IL8 and PCT showed exploratory associations with survival. Further studies are warranted to confirm our results.

Sepsis is a heterogeneous condition that causes millions of deaths each year [1]. It is defined as a dysregulated inflammatory response to an infection that leads to organ dysfunction [2]. Sepsis has been described as a time-dependent disease, since it has been shown that early bundle measures improve prognosis [3], however, tools to identify patients at risk of developing sepsis at an early stage remain limited.

In recent decades, significant advances in our understanding of the physiopathology of sepsis have generated several hypotheses, although clinical studies have failed to demonstrate the efficacy of these treatments [4]. Profiling inflammatory patterns may enable a more personalised approach reducing the risk of harm and increasing the likelihood of success [[5], [6], [7], [8]]. Experimental models and clinical observations have enabled us to understand how inflammatory mediators evolve in sepsis [9]. However, these have only been described in critically ill patients with sepsis [10] leaving the early temporal dynamics of cytokines in patients presenting to the emergency department insufficiently characterised.

As sepsis is a rapidly evolving syndrome in which the host response and organ dysfunction change markedly over time. A better understanding of the dynamics of cytokine/biomarker over time could help us to characterise sepsis more effectively [11]. Pro-inflammatory cytokines, the anti-inflammatory cytokine IL10, chemokines, IL1 receptor antagonist (IL1RN) and soluble TNF receptors (sTNFRs), as well as other inflammatory mediators or acute-phase reactants have been proposed as biomarkers for sepsis identification or prognosis for their physiopathology role in the development of the condition [9]. Whether early cytokine dynamics could be related to clinical events has not been adequately addressed in previous research. Dismissing the temporal dimension of biomarkers may lead to misclassification of patients’ true risk and underestimate their potential for guiding individualised therapeutic decisions.

We hypothesised that biomarker levels are associated with sepsis and the severity of organ dysfunction, resulting in different dynamics over time. Biomarker kinetics may also be useful for predicting survival.

Our aims were to analyse (1) the kinetics of inflammatory mediators over time in patients coming to the emergency department by infection during the first 24 h, (2) to understand whether sepsis might alter the trajectory of biomarkers, (3) whether the severity of sepsis, as measured by the SOFA score, were associated with biomarker values at different times, and (4) whether dynamic measures could aid survival prediction

Methods

Study design

This was a single-center, prospective, observational study conducted at Hospital Universitari Parc Taulí de Sabadell, in collaboration with the Institut d’Investigació i Innovació Parc Taulí (I3PT). Hospital Universitari Parc Taulí is a public university hospital that mostly treats patients from the Catalan regional health care system.

Study population

We included adult patients who presented to the Emergency Department (ED) with suspected infection. Patients were screened and enrolled consecutively during weekdays and periods when trained study staff were available. The emergency department attends approximately 50,000 medical emergencies per year. Inclusion was restricted to daytime periods, as the final sample was collected at 24 h, which could otherwise interfere with patients’ sleep.

Patients were included if they fulfilled the following criteria: age ≥18 years, suspected infection with a National Early Warning Score 2 (NEWS) score ≥3 [12], and willingness to provide written informed consent. Patients were excluded if they were pregnant or breastfeeding, receiving immunosuppressive therapy, antineoplastic chemotherapy, or chronic corticosteroids (prednisone ≥10 mg/day for ≥3 weeks), had received an organ transplant, had been included in another study within the previous 60 days, had been hospitalized for >48 h, had a life expectancy of less than 6 months, or had any limitation of therapeutic effort.

Study procedures

Eligible patients were assessed by ED physicians between September 2022 and June 2024.

Suspected infection was defined as the presence of clinical signs or symptoms compatible with an infectious process, as assessed by the attending emergency physician, together with the initiation of an infection-oriented diagnostic work-up and/or antimicrobial treatment according to routine clinical practice. The clinical assessment, including suspected infectious focus, vital signs, physical examination, laboratory findings, and the physician’s decision to manage the patient as having a possible infection. Upon obtaining informed consent, a 4 mL blood sample was scheduled at emergency department presentation, 4 h and 24 h after inclusion. Because sampling was performed in a real-world emergency department setting, exact sampling times could vary slightly from the nominal time points due to clinical care priorities and logistical constraints. Vital signs (heart rate, respiratory rate, blood pressure, oxygen saturation, body temperature, and level of consciousness) were recorded at 0 h, 4 h, and 24 h. Laboratory markers of inflammation and organ dysfunction, including differential white blood cell count, lactate, bilirubin, and creatinine, were measured at the same time points.

The final diagnosis of sepsis was determined according to Sepsis-3 criteria [2]: suspected infection and a ≥2 point increase in the Sequential Organ Failure Assessment (SOFA) [13] score, as confirmed independently by two blinded researchers (BC-AC). If there was no match, a consensus or a third researcher (AA) was used for the final diagnosis. Clinical assessment included evaluation of organ dysfunction (e.g., oliguria, altered mental status, respiratory distress), as well as sepsis scoring systems (qSOFA, NEWS), and microbiological data (e.g., blood, sputum, urine, skin, and stool cultures). Patients with a confirmed non-infectious diagnosis after discharge were excluded from the study.

Blood samples were centrifuged, and plasma was separated and stored at −80 °C. At the end of the study, samples were thawed and analysed for inflammatory mediators (IL1B, IL1RN, IL6, IL7, IL8, IL10, IL17A, TNF, TNF receptor superfamily member 1A (TNFRSF1A), interferon gamma (IFNG), CCL2 and CCL7) using a multiplexed bead-based immunoassay (Invitrogen PPX-12-MXRWGUZ), and PCT (FineTests ref. EH0341) and sTREM-1 (Invitrogen ref: EHTREM1) were analysed by ELISA methods. The study was purely observational and did not involve any interventions or modifications to standard clinical care. Patients were treated following local guidelines and the survival sepsis campaign recommendations [3,14,15].

The primary outcome was the development of sepsis, as defined above, and the secondary outcome was in-hospital mortality.

This study is reported following STROBE guidelines [16]. The institution’s Ethical Review Board approved the study (Comité de Ética de Investigación con medicamentos del Parc Taulí 2022/3004), which was performed following the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. Written informed consent was obtained from all patients or their legally authorised surrogates.

Statistical analysis

Descriptives

Continuous variables were summarised using means, medians, standard deviations, and interquartile ranges. Categorical variables were described as frequencies and percentages. Comparisons between groups were performed using the chi-square or Fisher’s exact test for categorical variables, and the Student’s t-test or Mann–Whitney U test for continuous variables, as appropriate.

Time-to-event data were analysed using Kaplan–Meier curves and compared using the log-rank or Gehan–Breslow–Wilcoxon tests.

Missing data and data transformations

For those measures below or above the limit of quantification, we imputed the minimum or maximum value.

Missing biomarker data were summarised at each time point. The primary longitudinal mixed-effects analyses were performed using available complete observations for each biomarker. Because missingness during follow-up could be related to clinical evolution, including death, early discharge, or loss to follow-up, we considered missingness to be potentially informative. Therefore, we performed a sensitivity analysis using a pattern-mixture modelling approach. Patients were classified according to their follow-up/missingness pattern as died, discharged, lost to follow-up, or reference/complete pattern. For each biomarker, we fitted mixed-effects models including time, missingness pattern, and the interaction between time and missingness pattern. These models were performed for the models including sepsis status and SOFA score. The aim of this sensitivity analysis was to assess whether estimated biomarker trajectories were materially affected by missingness patterns.

Due to the non-normal distribution of cytokine data, log10 + 1 transformation was applied before analysis.

Sample size

No formal prior sample size calculation was performed. As there was no existing data on the aims of this study, an exploratory approach was adopted.

Analyses of biomarker kinetics and associations to survival

Biomarker trajectories over time were analysed using linear and nonlinear mixed-effects models. The models included fixed effects for group (sepsis or non-sepsis at baseline-0 h), time (numeric, increasing with one unit each hour), and their interaction, and random effects as patient-specific intercepts and random slopes for time.

To assess the association between SOFA score and biomarker levels, linear mixed effects models were fit for biomarker levels (except for PCT and sTREM-1), with time (numeric) and SOFA score as fixed effects, and patient-specific intercepts and time slope as random effects. We also tested for interaction between SOFA score and time (p-value <0.05).

To allow for nonlinearity in the temporal profiles, time was modeled using a natural spline basis with boundary knots at 0 and 24 h and an internal knot at 4 h. Model fit was compared across alternative specifications using Akaike’s and Bayesian information criteria (AIC and BIC), and the best-performing time structure was retained for each biomarker. The NLME package [17] was used to fit the mixed-effects models.

To explore the predictive value of biomarker concentrations and their kinetics for in-hospital mortality, joint models for longitudinal and time-to-event data were used. These models combined a linear mixed-effects model for each biomarker with a relative risk model for survival, including age and maximum SOFA score during the first 24 h. The JMbayes2 package was used for joint modelling [18].

A Bayesian approach was used to fit the joint models, with posterior distributions approximated using three MCMC chains of 300,000 iterations each. The first 60,000 iterations were discarded as burn-in, and every 100th iteration was retained to reduce autocorrelation. Trace plots of the simulated values indicated good overlap between chains, suggesting stabilization. Convergence to the posterior distribution was further assessed using the potential scale reduction factor (R̂), with all values close to 1, indicating convergence. We used the default prior distributions implemented in JMbayes2 [18], which correspond to weakly informative priors for fixed effects, variance components, and the association structure between the longitudinal and survival submodels.

R version 4.4.0 was used for all analyses [19].

Results

Between September 2022 and June 2024, 214 patients were screened and accepted as participants. Of these, 196 satisfied the inclusion/exclusion criteria and 173 had data available for analysis (at least one measure of each biomarker/inflammatory mediator), as well as having their survival recorded at discharge (see Fig. 1). 137 patients (79%) developed sepsis within 24 h of arriving at the ED and the rest were categorized as infections without sepsis n = 36 (21%). The researchers reached a final diagnosis agreement in 97% of cases (n = 168); the remaining 5 cases were solved by consensus; Kappa Cohen index 0.90 CI95% 0.82; 0.98.

Fig. 1.

Flowchart.

Study recruitment overview. Adjudication of sepsis was performed by at least two physicians based on the Sepsis-3 definition [2].

Baseline characteristics and comorbidities are summarised in Table 1. Patients who developed sepsis were older and had similar comorbidities to those who did not. The sites of infection were mostly respiratory and urinary. The most commonly isolated pathogens (bacteremic or non-bacteremic) in non-septic patients were Escherichia coli and Klebsiella pneumoniae, whereas in septic patients, the most commonly isolated pathogens were Escherichia coli and Streptococcus pneumoniae (eTable 1). The outcomes for patients with sepsis were worse, requiring more ICU admissions and significantly lower survival rates with an in-hospital mortality of 29 (21%) in the group of sepsis compared with none in the non-septic group (see eFig. 1).

Table 1.

Demographic and clinical characteristics at time 0 of patients by sepsis status.

Variable	Non-Sepsis (n = 36)	Sepsis (n = 137)	p-Value
Age, years (median [IQR])	69.5 [55.75, 79]	78 [67.00, 85]	0.005
BMI, kg/m2 (median [IQR])	26.6 [24.2, 31.3]	25.75 [22.52, 29.62]	0.40
Sex male, n (%)	29 (80.6%)	87 (63.5%)	0.08
Charlson comorbidity index (median [IQR])	5 [3,7]	5 [4,7]	NS
Coronary disease, n (%)	4 (11.1%)	18 (13.1%)	1.0
Chronic heart failure, n (%)	4 (11.1%)	28 (20.4%)	0.29
Peripheral vascular disease, n (%)	5 (13.9%)	19 (13.9%)	1.0
Cerebrovascular disease, n (%)	5 (13.9%)	29 (21.2%)	0.45
Chronic pulmonary disease, n (%)	13 (36.1%)	36 (26.3%)	0.33
Dementia, n (%)	6 (16.7%)	34 (24.8%)	0.41
Liver disease, n (%)	4 (11.1%)	8 (5.8%)	0.27
Diabetes mellitus, n (%)	8 (22.2%)	39 (28.5%)	0.58
Chronic kidney disease, n (%)	10 (27.8%)	29 (21.2%)	0.53
Site of infection			0.11
Respiratory	21 (60%)	55 (40%)
Urinary	8 (23%)	38 (28%)
Abdominal (non-surgical)	1 (3%)	25 (18%)
Other	5 (13%)	17 (13%)
Days since the onset of symptoms, Median [Q1, Q3]	2 [0,3]	2 [0,4]	0.99
Systolic blood pressure mmhg (median [IQR])	114.5 [94.75, 130.75]	97 [85,114]	0.007
Diastolic blood pressure mmHg (median [IQR])	64.5 [57.25, 82.25]	57 [49, 68]	0.003
Respiratory rate per min (median [IQR])	34 [31, 36.50]	23.50 [20, 30]	0.021
Heart rate per min (median [IQR])	110 [86.5, 122.25]	98 [78, 115.50]	0.07
SOFA Score (median [IQR])	1 [1,1]	3 [2,5]	<0.0001
APACHE Score (median [IQR])	9 [6.75, 11]	12 [9.50, 16]	0.0005
NEWS2 Score (median [IQR])	5 [4,6]	6 [4,8]	0.06
qSOFA Score (median [IQR])	1 [0,1]	1 [0.75, 1]	0.0007
SpO2/FiO2 ratio (median [IQR])	442.86 [383.33, 457.14]	419.05 [339.29, 452.38]	0.21
PaO2/FiO2 ratio (median [IQR])	333.33 [264.52, 380.95]	277.38 [196.62, 380.95]	0.32
Leukocytes 109/L(median [IQR])	12.34 [9.55, 15.85]	12.36 [7.52, 18.34]	0.79
Lymphocytes % (median [IQR])	8.00 [4, 12.5]	6.00 [4, 10.25]	0.79
Neutrophils % (median [IQR])	85 [70.5, 89.5]	87 [79.75, 91]	0.21
Absolute neutrophils/lymphocytes ratio	10.6 [7.25, 20.4]	14.1 [7.17, 19.6]	0.616
CRP mg/dL(median [IQR])	6.28 [4.03, 16.12]	11.69 [3.31, 25.84]	0.11
Lactate mmol/L (median [IQR])	1.68 [1.31, 2.22]	2.51 [1.49, 3.54]	0.005
pH (median [IQR])	7.38 [7.34, 7.43]	7.38 [7.33, 7.43]	0.81
PaO2 mmHg (median [IQR])	80.00 [75.00, 82.00]	80.00 [64.38, 87.92]	0.83
PaCO2 mmHg (median [IQR])	41.00 [37.28, 44.65]	35.20 [32.00, 43.20]	0.0445
Bicarbonate meq/L (median [IQR])	23.20 [22.50, 26.10]	21.90 [19.40, 24.58]	0.0326
Creatinine mg/dL (median [IQR])	0.94 [0.82, 1.11]	1.24 [0.95, 1.97]	0.0015
Glucose mg/dL (median [IQR])	125 [106.50, 145.25]	143 [118, 185.50]	0.051
Bilirubin mg/dL (median [IQR])	0.6 [0.4, 0.75]	0.8 [0.5, 1.35]	0.0028
Bacteremic disease	7 (19.4%)	35 (25.5%)	0.0405
Adequate antibiotics, n (%)	29 (80.6%)	118 (86.1%)	1.0
ICU admission, n (%)	1 (2.8%)	29 (21.2%)	0.019
Septic shock, n (%)	0 (0.0%)	19 (13.9%)	0.014
Mortality, n (%)	0 (0.0%)	29 (21.2%)	0.005

APACHE, Acute Physiology and Chronic Health Evaluation; BMI, body mass index; CRP, C-reactive protein; FiO₂, fraction of inspired oxygen; ICU, Intensive Care Unit; IQR, interquartile range; MV, mechanical ventilation; NEWS, national early warning score; PaCO₂, arterial partial pressure of carbon dioxide; PaO₂, partial pressure of arterial oxygen; qSOFA: quick sequential organ failure assessment score; SpO2: peripheral oxygen saturation SOFA: sequential organ failure assessment score.

A swimmer-plot summarising the description of samples obtained and follow-up is provided in eFig. 2. Missing biomarker data at each time point are summarised in eTable 2.

Multiplex bead analyses for inflammatory mediators were performed in duplicate, and the quality and reading description are provided in eTable 3. The trajectories of each biomarker are shown in eFig. 3 and raw distributions in eFig. 4.

Dynamics of biomarkers according to sepsis diagnosis

Linear and nonlinear mixed models were used to evaluate the associations between biomarkers over time and sepsis diagnosis at baseline. Biomarkers with more than 50% of measurements below good quality were excluded from the longitudinal analysis (IFNG, IL1B, IL7, IL17A, CCL7, TNF) (eTable 3). The models’ fit was evaluated by AIC and BIC (eTable 4). Only the nonlinear model evaluating IL6 showed better AIC and BIC (eTable 4) than linear model using a natural spline with boundaries 0, 24 h and a knot at 4 h. In all patients with suspected infection included, IL10 (β estimate CI95% −0.016, −0.028 to −0.004), IL1RN (β estimate CI95% −0.014, −0.024 to −0.004) and IL6 (β estimate for the first time spline 0−4 h CI95% −1.021, −1.609 to −0.433) showed a significant decrease over time (per hour). Sepsis diagnosis was associated with higher values of IL1RN (β estimate, 95% CI: 0.378, 0.153–0.603) and TNFRSF1A (β estimate, 95% CI: 0.40, 0.21–0.58). Only IL6 showed a significant interaction between a sepsis diagnosis and time for the second-time spline (4–24 h) (β estimate CI95% −0.330, −0.653 to −0.007) (Fig. 2). The results are summarised in eTable 3 and eFig. 4.

Fig. 2.

IL6 trajectories over time by sepsis diagnosis using a nonlinear mixed-effects model.

Trajectories of plasma concentration IL6 during the first 24 h by groups assigned whether or not the patients developed sepsis. Lines represent the fitted linear mixed model. Each point represents the concentration of IL-6 for each subject. The Y-axes are represented on a log scale.

Association of the severity of SOFA score on biomarker levels

IL10 (β CI95% 0.048, 0.021–0.075), IL1RN (β CI95% 0.044, 0.017–0.071), CCL2 (β CI95% 0.046, 0.021–0.071), TNFRSF1A (β CI95% 0.050, 0.030–0.070) and PCT (β CI95% 2.15 1.26–3.03) showed a significant association with SOFA score, where higher biomarker concentrations coincide with higher SOFA score. There was also a significant association between the time in hours spent in the hospital and the levels of the same biomarkers, except for PCT (eTable 6, eFig. 5). Only the interaction between time and SOFA score was significant for IL6, showing that IL6 increased during the first spline (0–4 h) with higher SOFA score (β CI95% 0.122, 0.006–0.238) and decreased during the second spline (4–24 h) with higher SOFA score at 24 h (β CI95% −0.080, −0.133 to −0.027) (Fig. 3). The results are summarised in eTable 6 and eFig. 5. A linear regression model between IL6 delta and SOFA score delta (24-h results minus baseline results) was performed to corroborate whether patients with a higher fall in IL6 levels have a higher fall in SOFA score. No correlation was observed between the deltas (Fig. 3).

Fig. 3.

Nonlinear mixed-effect plot for IL6, and SOFA score according to time sampling and relation between changes of IL6 and SOFA score at 24 h.

Above: Linear mixed model estimations of plasma concentration of IL6 according to SOFA score and time sampling. Lines represent the fitted linear mixed model. Each point represents the concentration of IL6 for each subject. The Y-axes are represented on a log scale. Below: Changes in SOFA score between 0 and 24 h associated with changes in log-transformed IL6 levels over the same period, with the regression line and corresponding slope, correlation coefficient, and coefficient of determination indicating the strength and direction of this relationship. β = regression coefficient. r = Pearson correlation coefficient. R² = coefficient of determination. p = p-value.

Sensitivity analyses

Because missingness during follow-up could be related to clinical evolution, including death, early discharge or loss to follow-up, we performed a pattern-mixture sensitivity analysis. Overall, the results were consistent with the primary longitudinal mixed-effects models. For most biomarkers, the interaction between time and missingness pattern was not statistically significant, suggesting that the estimated temporal slopes were not materially driven by the missing-data pattern. The main longitudinal findings were preserved but some differences in baseline biomarker levels were observed according to missingness pattern, and PCT showed a significant time-by-lost-follow-up interaction. However, this finding was limited to the lost-to-follow-up pattern and did not modify the overall interpretation of the longitudinal biomarker analyses. Full results of the pattern-mixture models are provided in eTables 7 and 8.

To account for left censored values by LLQ imputation and log10 + 1 transformation we performed the following analyses. Among biomarkers, IL10 had a considerable proportion of LLQ values, with approximately 24% extrapolated below the LLQ and 12% set to the LLQ. Therefore, we performed a censored mixed-effects sensitivity analysis for IL10 (eTable 9). The IL10 decrease over time remained consistent with the primary mixed-effects model, indicating that the observed temporal trajectory was not explained only by lower-limit substitution or by the log10(x+1) transformation. A supplementary figure showing raw values below 10 units was added to illustrate the distribution of observations in the range most affected by the additive offset (eFig. 4).

Joint models for longitudinal and time-to-event data

Associations between the longitudinal concentrations of inflammatory biomarkers and in-hospital time-to-death were evaluated in joint model analyses. IL8, TNFRSF1A, and PCT were associated with the hazard of in-hospital death (Table 2).

Table 2.

Results of joint modelling for longitudinal measures and time to event.

Parameter of interest	Parameter	Hazard ratio	Lower 95% Credible Interval	Upper 95% Credible Interval	PBayes
Joint modelling evaluating IL8	Age	1.068	0.955	1.182	0.202
	Highest SOFA score	1.415	1.025	1.856	0.042
	IL8	0.655	0.582	0.728	0.000
Joint modelling evaluating TNFRSF1A	Age	1.066	0.948	1.181	0.216
	Highest SOFA score	1.444	1.042	1.892	0.035
	TNFRSF1A	0.505	0.419	0.682	0.010
Joint modelling evaluating PCT	Age	1.085	1.032	1.145	0.002
	Highest SOFA score	1.402	1.174	1.673	0.001
	PCT	1.004	1.001	1.008	0.000

Results are obtained by joint models analyses, corrected for age, and the highest SOFA at the day of sampling. Joint model analysis was used to evaluate the longitudinal changes in plasma biomarker concentrations and the association between changes in biomarker concentration and outcomes. Joint modelling analyses two types of data: repeated measurements over time and time-to-event data (survival data). P_Bayes indicates the posterior probability that the parameter differs from zero, as estimated from the joint Bayesian models. Credible intervals represent the ranges containing the true parameter value with 95% posterior probability.

Results for the remaining biomarkers evaluated in the joint models (IL10, IL1RN, IL6, CCL2, sTREM-1), for which no association with in-hospital mortality was identified, are presented in eTable 10.

The joint models were tested for convergence, discrimination and calibration. The discriminatory capacity was evaluated using receiver operating characteristic (ROC) curves, which were constructed to assess the ability to predict the event at 1024 h based on biomarkers measured at 24 h. The area under the curve for the ROC curves for each model was as follows: 0.56 for evaluating the value of log TNFRSF1A, 0.63 for evaluating the value of log IL8, and 0.60 for evaluating PCT. The Brier scores were 0.13, 0.12 and 0.16, respectively. The ROC curves and calibration plots are shown in eFigs. 7–8.

Discussion

In this study, we analysed the dynamics of inflammatory markers in patients suspected of having an infection within the first 24 h of receiving ED care. Our main finding comprises: IL10, IL1RN, and IL6 decreased in the first 24 h in all included patients, while IL1RN increased with the diagnosis of sepsis. Only IL6 showed a significant change in trend when patients had a sepsis diagnosis; IL6 had a fast step decrease in those patients after the first 4 h (Fig. 2). For the association between the SOFA score and the inflammatory mediators over time, IL10, IL1RN, CCL2, TNFRSF1A, and PCT increased along with the SOFA score. Time measured in hours was associated with decreased values, with the exception of PCT, meaning that the biomarker value may be associated with a lower SOFA score on arrival or a higher SOFA score after 24 h. In the case of IL6, the interaction between time and SOFA score was significant, confirming the findings of a sepsis diagnosis, meaning that the association between SOFA score and IL6 was reversed after the first 4 h (Fig. 3). These findings highlight the importance of accounting for time when interpreting inflammatory mediator concentrations. They also suggest that the relationship between inflammation and organ dysfunction may differ across the early phases of the host response. Regarding joint models results, time-dependent concentrations of TNFRSF1A, IL8 and PCT during the first 24 h were associated with survival probability. However, the discriminatory performance of the models was modest, and the directions of the associations for TNFRSF1A and IL8 were unexpected. These findings should therefore be regarded as hypothesis-generating and require validation in independent cohorts.

The importance of our findings lies in the fact that, despite advances in the pathophysiology of sepsis, few clinical studies have focused on the kinetics of inflammatory mediators during the earliest hours of hospital care. Longitudinal analysis of data enables us to understand the behaviour of inflammatory responses and to characterise patients dynamically. Van Deuren described the kinetics of TNF, TNFRSF1A, IL1B, and IL1RN. He based his observations mainly on those in critical care patients. While IL1B and TNF showed a decrease within the first 8 h, TNFRSF1A and IL1RN exhibited greater variability, demonstrating a smaller decrease within the first 8 h [10]. In another study describing the evolution of meningococcal disease, TNF and IL1B were below the limit of quantification, as in our study, and although IL6 levels increased, ex vivo production was suppressed [20]. A study describing the kinetics of IL6 after septic shock onset found that the IL6 peak occurred near the shock onset and decreased within 24 h [21]. The kinetics of IL6 showed that a rapid decline within 24–48 h was associated with successful antimicrobial therapy [22]. In our study, patients with sepsis or worst SOFA score had a significantly higher probability of a steeper decrease in IL6 levels over time (per hour). However, these analyses should be interpreted as exploratory and hypothesis-generating because of the limited sample size, particularly for interaction estimates.

PCT is one of the widely studied biomarkers in sepsis and has been shown to be associated with prognosis [[23], [24], [25], [26]] and the capacity to guide antibiotic stewardship [[27], [28], [29]]. PCT may increase in the first 24 h [30], followed by a decrease in the next days. Patients who did not achieve a decrease in the levels of more than 80% have a lower survival probability [30]. Consistent with this, we observed an exploratory association between longitudinal PCT concentrations and survival probability.

Kellum et al. [31], describing the kinetics of IL10, TNF and IL6 in a cohort of patients with pneumonia, during the first days of admission. They observed, as in our study, that there is an overlap in biomarker levels between those who developed or did not develop sepsis (former severe sepsis) (Fig. 2 and eFig3), and when categorising the continuous variables, mortality was higher in those with high levels of inflammatory mediators. We observed that patients with a confirmed infection and a clear systemic response (NEWS2 > 3), but not sepsis, may exhibit a similar dysfunctional immune response in terms of biomarkers as those with sepsis. This emphasises the need for a clearer definition of dysregulated responses and the concept of pre-sepsis [32]. Surprisingly, the IL8 and TNFRSF1A were associated with an HR below one, indicating that those with higher values in repeated measures show a better prognosis. This is opposed to what has been observed in other studies with single measures. A hypothesis that may explain these results could be that persistently elevated levels may characterise patients with a sustained and regulated inflammatory response, whereas fluctuating values may reflect most inflammatory dysregulation. Nevertheless, given the exploratory nature of this finding and its inconsistency with previous literature, these results should be interpreted cautiously and confirmed in further studies.

In recent years, several endotypes or subphenotypes have been identified in sepsis and other critical illnesses, depending on the patient’s inflammatory status or genomic expression [[33], [34], [35]]. Nevertheless, inflammation otherwise to some chronic conditions is not static in sepsis, changing faster even in a short time. While some retrospective analyses of clinical trial data have demonstrated the potential for improved management based on different subphenotypes or profiles [17,18], the heterogeneity of patients and the observation that subphenotypes can change within the first few hours or days underscore the necessity for a more comprehensive understanding of the temporal evolution of inflammatory mediators [[36], [37], [38], [39]].

Our study has strengths and limitations that should be acknowledged. The principal strength and contribution of this study is the characterisation of the early kinetics of multiple inflammatory mediators during the first 24 h after Emergency Department presentation with suspected infection. Additional strengths include the collection of repeated samples, the adjudication of sepsis status by blinded researchers, and the use of mixed-effects models for repeated measurements [40] and joint models for longitudinal and time-to-event data. These methods are the most appropriate approach for assessing the impact of potential biomarkers and for dynamically updating patient prognosis [41]. We chose three times, considering the baseline, the response to early bundles and the response to 24 h. This study has also several limitations that should be acknowledged. First, its single-center design may limit the generalizability of the findings to other settings and patient populations. Second, the relatively small size of the non-sepsis group may have reduced the statistical power to identify differences between groups. Recruitment was restricted to periods when trained study personnel were available, and we were unable to quantify the number or characteristics of potentially eligible patients who were not approached, which may have introduced selection bias. In addition, no prespecified sample-size calculation was performed and patients in whom infection was ruled out after enrollment were excluded from the analysis. Finally, the evaluation of multiple biomarkers and their longitudinal associations with clinical outcomes was exploratory, increasing the possibility of chance findings due to multiple comparisons.

In conclusion, our comprehensive study has improved our understanding of how sepsis and organ dysfunction could be associated with changes on inflammatory mediators/markers during the first 24 h after hospital admission for infection. We found that IL6 exhibits exploratory evidence of different kinetics depending on whether or not the patient develops sepsis or higher SOFA score within the first 24 h. IL6, IL10, IL1RN, CCL2, TNFRSF1A and PCT values were associated with the SOFA score, with different values observed at each time point (except for PCT). Finally, we observed an association between survival and the longitudinal data of TNFRSF1A, IL8 and PCT. However, the discriminatory capacity of these associations was low, and the direction of the associations between IL8 and TNFRSF1A requires confirmation in further studies.

CRediT authorship contribution statement

AC & AA: Study supervision. BC, VM, MCR, AAB, ECD, EG, LB, AA and AC participated in protocol development, study design, study management, and data interpretation. AF, LK, RvA, LB y AC participated in statistical analysis and data interpretation. BC, ACa, EN, PS, CG, JT, GG, MLM, IM, AT, MV, MC, MS, XB, MLR, TT, MFS, CS, IuM participated in patient inclusion and data collection. VM, MCR, AAB and ECD participated in sample analyses.

All authors read and approved the final manuscript.

Ethical statement

The institution’s Ethical Review Board approved the study (Comité de Ética de Investigación con medicamentos del Parc Taulí 2022/3004), which was conducted in accordance with the ethical standards set out in the 1964 Declaration of Helsinki and its subsequent amendments. Written informed consent was obtained from all patients or their legally authorised surrogates.

Declaration of Generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the author(s) used DeepL Write to refine language, review English grammar and improve clarity. After using this tool, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.

Funding

This study was supported by the Ministry of Science: CPP2021-008394. CIBERES, and Research and Innovation Institute Parc Tauli (I3PT-CERCA). AC acknowledges receiving financial support from Instituto de Salud Carlos III CD21/00087, MV23/00097 and PI 23:00106. The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and in the decision to submit the manuscript for publication.

Data availability

The datasets used and/or analysed during the current study are available from CORA.Repositori de Dades de Recerca, https://doi.org/10.34810/data3142.

Declaration of competing interest

The authors have no conflict of interest.

Appendix A. Supplementary data

The following is Supplementary data to this article: