In this study, we evaluated the plasma protein concentrations of 58 biomarkers in 20 pediatric sepsis patients admitted to the PICU and compared their biomarker levels to those of 20 age- and sex-matched healthy control subjects. Multiple plasma biomarkers were significantly increased in sepsis patients, with nine biomarkers yielding AUC values of > 0.9 on their respective ROC curves. Moreover, combinations of these biomarkers yielded AUC values of 1.00, with five combinations reaching this parameter: IL-8 & IL-6; MCP-1 & HSP70; MCP-1 & IL-6; HSP70 & Hyaluronan; and Hyaluronan & IL-6. Lastly, many biomarkers were correlated with clinical variables and severity of illness scores, suggesting prognostic utility.

Our patients were similar to other pediatric sepsis studies in demographics, illness severity, and clinical interventions [18, 19]. The inciting pathogens were variable, including bacterial, viral, and fungal, with a primary respiratory source. Systemic inflammation and immune dysfunction are characteristic features of pediatric sepsis [20]. A pro-inflammatory stage occurs early, characterized by the recruitment of leukocytes to infected tissues [21]. Once activated, the recruited cells secrete additional pro-inflammatory cytokines [22], chemokines [23], and reactive oxygen species [24]. Local accumulation of these proinflammatory mediators likely instigates microvascular endothelial injury and glycocalyx shedding (e.g., hyaluronan) [25, 26].

Conventional statistical techniques identified 24 plasma biomarkers that were significantly different in sepsis patients from controls. The majority of elevated biomarkers were either pro-inflammatory cytokines [27, 28] or MMPs [29], zinc-containing endopeptidases capable of degrading extracellular matrix proteins and processing bioactive molecules [30]. While 21 biomarkers increased in sepsis, 3 biomarkers decreased from control levels (RANTES, MMP9, and MMP2). Decreased RANTES (or CCL5), a chemokine that regulates CD8 T-cell responses during viral infection, was also depressed from baseline levels in both neonatal and pediatric sepsis [28, 31]. The expression of several plasma MMPs is altered by sepsis in adults, with elevated MMP9 and depressed MMP2 [32]. The depressed MMP9 measured in pediatric sepsis [33] may represent antibody specificity to latent versus active forms or reflect age-dependent differential sepsis responses.

Using machine learning, six biomarkers (MCP-1, M-CSF, IL-6, MMP3, MMP10, and MMP7) were identified that shared high importance between both groups (PICU Day-1 and PICU Day-2) and may be useful for future studies evaluating biomarkers in pediatric sepsis patients who present later in their course of illness than those captured within this study.

The five biomarker combinations that yielded AUC values of 1.00, are individually known to be associated with sepsis and inflammation. IL-6 is an early pro-inflammatory cytokine that induces the synthesis of acute-phase proteins, stimulates antibody production, affects T-cell development, and promotes the differentiation of non-immune cells [34]. IL-8 and MCP-1 are chemokines that recruit leukocytes [35, 36]. Hyaluronan is a constituent of the microvascular glycocalyx [26] and, in response to inflammatory cytokines, is catalyzed into fragments that promote further inflammation [37]. Unlike proteins involved in inflammatory cascades, HSP70 is released by active secretion or during lysis by necrotic cell death, and its levels are often variable in oxidant states. In addition, HSP70 levels correlate with sepsis mortality [38].

We identified multiple sepsis biomarkers that correlated with clinically significant variables. MMP7, an enzyme that degrades the extracellular matrix [32], was negatively correlated with age and male sex, making this biomarker more likely to be present in pediatric sepsis patients who are young and female. MMP8, a neutrophil collagenase, is positively correlated with a respiratory source of infection [39]. The presence of MMP2 is negatively correlated with having an abnormal finding on a chest x-ray, as well as the need for invasive mechanical ventilation [40]. These findings suggest that patients with higher levels of this biomarker are less likely to have an abnormal chest x-ray and less likely to require invasive mechanical ventilation while hospitalized. Additionally, the presence of HSP70 was negatively correlated with the need for inotropic medication, suggesting that patients with increased HSP70 will have overall greater hemodynamic stability throughout their illness course [41].

Multiple biomarkers infer utility in predicting illness severity when matched to established illness severity scores. The biomarkers with the strongest illness severity correlations were G-CSF, M-CSF, and IL-8. These three biomarkers were each positively correlated with PRISM III scores, initial PELOD-2 scores, and the highest measured PELOD-2 scores. In keeping with their role as sepsis severity biomarkers, each of these biomarkers negatively correlated with GCS on admission. As for other markers of sepsis severity, MMP1, MMP3 [42], and MMP10 were all positively correlated with PIM-2, a score that predicts the risk of death for pediatric patients admitted to intensive care. Given biomarker changes that significantly correlate with illness severity, these biomarkers might be useful for sepsis prognosis.

Biomarkers correlating with PICU length of stay and overall hospital length of stay were also identified. Elastase 2, NGAL, and MIG positively correlated with PICU length of stay, whereas hyaluronan and MIG positively correlated with hospital length of stay. Elastase 2 and NGAL are released by activated neutrophils, MIG is a T-cell lymphocyte chemokine, and hyaluronan is cleaved from the microvascular endothelial glycocalyx. These biomarkers may represent a severe host-pathogen response as well as a prolonged recovery process.

Importantly, IL-3 was noted to be increased in pediatric patients who had a viral agent identified rather than a bacterial infection. IL-3 is produced by activated T-cells and increases innate antiviral immunity by promoting the recruitment of circulating plasma dendritic cells [43]. IL-3 may help distinguish which sepsis patients require antibiotic versus antiviral coverage and aid antimicrobial stewardship.

The importance of this study is highlighted by the number of blood biomarkers measured, together with advanced analytics. Indeed, recent pediatric sepsis biomarker studies focused on only a few biomarkers analyzed with conventional statistics [44,45,46,47]. The machine learning incorporated herein identified the biomarker importance for determining sepsis. We have also taken a novel approach to identify combinations of biomarkers that, when used together, yield potential utility in predicting sepsis severity.

The clinical relevance of this study is two-fold: first, identification of biomarker combinations may aid earlier recognition of sepsis in children; and second, the biomarker(s) that correlate with clinical variables and illness severity may aid prognostication. Quantitative laboratory and semi-quantitative point-of-care immunoassays could be constructed to measure multiple biomarkers identified here for early sepsis diagnosis, including multiplex, lateral flow, and chip assays. Future studies, using targeted proteomics and machine learning, should measure the plasma proteome in pediatric sepsis to identify novel biomarkers not yet considered, as has been done for adult sepsis [48].

Our study identified combinations of plasma biomarkers that act as significant markers of sepsis. Individually, the identified sepsis biomarkers share commonalities with other inflammatory conditions, including diabetic ketoacidosis [49, 50], trauma [51], and ischemic injury [52]. In contrast, biomarker combinations hold value for disease specificity and may allow for the distinction of sepsis from other inflammatory states, such as those listed above. Nonetheless, several limitations of our study are worthy of discussion. First, our study includes a limited sample size and was conducted at a single center. Despite this, the data yielded robust results even while using conservative statistical methods. Second, the pathogens were variable, with the majority suffering from bacterial illness as compared to viral pathogens [53]. Nonetheless, bacterial infections are amenable to antibiotic treatment, and early recognition and intervention are critical to improving outcomes [53]. Also, IL-3 seemed to distinguish bacterial from viral infections, thereby requiring further study. Despite these caveats, our exploratory data are important for future hypothesis-driven studies in larger populations.