This is a secondary analysis of a recently published study (19). Ethical approval was obtained from the Human Studies Protection Office of the Penn State College of Medicine (#15328, 7/30/2020). Procedures were followed in accordance with ethical standards established by the Human Subjects Protection Office and with the Helsinki Declaration of 1975. The Modified Early Warning Scoring (MEWS) scoring-based algorithm was used to identify critically ill patients with possible sepsis (20, 21) from November 2021 to August 2023. Dual, independent investigator reviews ensured unbiased selection of patients meeting inclusion criteria from electronically flagged patient records. Informed consent was obtained from patients having decision-making capacity, or from the legally authorized healthcare representatives of patients lacking decision-making capacity.

Eligible septic candidates were consecutively identified adults, >18 years, within 48 hours of critical illness onset. Sepsis was defined by Sepsis-3 criteria, namely a change in SOFA score of two or more in the setting of clinically suspected or microbiologically proven infection (22). Non-survivors were defined as those deceased within 30 days post-enrollment. Patients exhibiting active hematologic malignancies and those receiving immune-altering therapies were excluded.

Clinical data, including whole blood counts and cell differentials, were retrieved from hospital records and post-discharge interviews. Illness severity was measured using Charlson Comorbidity Index (23) and APACHE II score (24–26). Outcomes included organ dysfunction via SOFA score; 30-day mortality, readmission, and infection rates. Secondary infection was defined using the CDC/NHSN definitions (27). Blood samples were collected early, within 48h of ICU admission, for leukocyte analysis, cytokine analysis, and cytokine quantification after ex vivo stimulation tests, with supernatants frozen for later analysis.

Whole blood was collected in sodium heparin tubes and ex vivo cytokine production was performed on the day of collection. Plasma was separated by centrifugation (400 g x 5 min). Plasma was aliquoted and stored at -80°C until subsequent batch analysis. Cytokines were measured using the Ella™ microfluidic automated immunoassay system (Bio-Techne, MN). Biomarkers were measured by batch analysis by Ella™ using custom manufactured 32 sample x 4 analyte cartridges (28–45) ( Supplementary Table 1 ). Plasma samples were thawed and diluted 1:10 with sample diluent 13, and 50 µL of diluted sample were loaded into each sample inlet, and data was acquired utilizing the Simple Plex Runner software v.3.7.2.0 (ProteinSimple).

Whole blood was collected in sodium heparin tubes. As previously described (19) 50 μL whole blood was diluted ten-fold in HEPES-buffered RPMI media, then exposed to designated stimulants (46, 47). Whole blood samples from each study participant were exposed to separate conditions: (1) 500 pg/mL lipopolysaccharide (LPS) from Salmonella enterica strain abortus equi, or (2) 10 ng/mL phorbol 12-myristate 13-acetate (PMA) with 1 μg/mL ionomycin. Following incubation for 4 hours at 5% carbon dioxide and 37 °C, and subsequent centrifugation, supernatants were frozen at -80 °C until the time of analysis. The Ella™ automated immunoassay system (Bio-Techne, Minneapolis, MN) was used for triplicate measurement of interferon-gamma (IFN-γ), tumor necrosis factor (TNF), and interleukin (IL)-6. Data was acquired utilizing the Simple Plex Runner software v.3.7.2.0 (ProteinSimple).

Multicolor flow cytometry was performed on a 17-color Becton Dickinson (BD) FACSSymphony. Blood was obtained from three separate healthy adult controls, ages 40-50, on the morning of stimulation. After 3h of ex vivo stimulation, brefeldin A was added to cause cytokine accumulation in cells. After 4h whole blood cells were surface stained with fluorescent-labeled antibodies (BD or BioLegend) and intracellular staining for TNF and IFN-γ was performed using Cytofix/Cytoperm (BD) kit.

Using Prism (v9.5) and JMP^® Pro (v18.0.2), with a significance threshold of 0.05, we summarized variables descriptively. Continuous variables’ distributions were evaluated via histograms, probability plots, and normality tests. Demographic comparisons between groups utilized Chi-square and two-sample t-tests. Due to the non-normal distribution of the cytokine data, log₁₀ transformation was applied before modeling. Designation of the inflammatory group (hypo-inflammatory vs. hyper-inflammatory) was determined using a parsimonious mathematical model with IL-6 (pg/mL), TNFR1 (pg/mL) and bicarbonate (mEq/L) values as variables that was previously validated for patients with ARDS (7).

In the primary study, we enrolled a total of 60 individuals with sepsis. Detailed demographic profiles and clinical outcomes of these patients have been previously described (19). Multiple cytokines were measured in sepsis patients using custom Ella™ cartridges. Using a previously validated parsimonious classifier model (7) based on IL-6, TNFR1 and bicarbonate levels, subjects were divided into two clusters: hyper-inflammatory (defined by elevated IL-6, TNFR1 and low bicarbonate levels, n=18, red), and hypo-inflammatory (n=42, blue) groups ( Figure 1 ). The classifier model is a logistic regression-based model comprised of IL-6, sodium bicarbonate and TNFR1. For the model, IL-6 and TNFR1 were log-transformed after the addition of one (+1). The coefficients used for the model were detailed in previous manuscripts (7, 48). We used the validated cutoff of 0.5 for this analysis.

Not surprisingly, TNFR1, IL-8 and IL-6 levels were significantly (***p<0.001) different between the inflammatory subphenotypes ( Figure 1A ). Also not surprisingly, given that bicarbonate is part of the validated model, bicarbonate levels were significantly lower in the hyper-inflammatory cluster ( Figure 1B ). 30-day mortality was almost four-fold higher (p=0.0046) in the hyper-inflammatory (44.4%) group than the hypo-inflammatory (11.9%) cluster ( Figure 1B ).

We investigated the “functional immune response” by quantifying the amount of cytokines produced by whole blood in response to precise amounts of immune stimulation. We quantified total TNF production after 4h of ex vivo stimulation of fresh whole blood with LPS. Interestingly, patients from the hyper-inflammatory cluster demonstrated lower ex vivo TNF production as compared to the hypo-inflammatory group (407±442 vs 729±637 pg/mL, p = 0.0159 ( Figure 2A ). As a side experiment, we investigated what cell types were producing TNF in response to LPS stimulation using intracellular cytokine staining (ICS)( Supplementary Figure 1 ). Patients with sepsis had reduced TNF production as compared to healthy controls ( Supplementary Figures 2 , 4 ). While CD15+ neutrophils produced some TNF, by far the largest producers of TNF in the ex vivo assay are CD14+ monocytes ( Supplementary Figures 2–4 ).

To control for the potential impact of leukopenia on TNF production, we normalized TNF levels to monocyte counts calculated using daily complete blood count with automated differential cell count analysis. TNF production per monocyte (pg TNF/10⁶ cells) did not differ between the inflammatory clusters. In contrast to TNF production in response to LPS stimulation, the total amount of IFN-γ production (pg/mL) following PMA/ionomycin stimulation did not differ between the hyper- and hypo-inflammatory groups ( Figure 2C ), nor was there a difference after normalizing to lymphocyte count (pg IFN-γ/10⁶ cells) ( Figure 2D ).