Identification of Clinically Significant Cytokine Signature Clusters in Patients with Septic Shock

Abstract

Objectives:

To identify cytokine signature clusters in patients with septic shock.

Design:

Prospective observational cohort study.

Setting:

Single academic center in the United States.

Patients:

Adult (≥ 18 year old) patients admitted to the Medical Intensive Care Unit with septic shock requiring vasoactive medication support.

Interventions:

None

Measurements and Main Results:

114 patients with septic shock completed cytokine measurement at time of enrollment (t₁) and 24 hours later (t₂). Unsupervised random forest analysis of the change in cytokines over time, defined as delta (t₂-t₁), identified three clusters with distinct cytokine profiles. Patients in Cluster 1 had the lowest initial levels of circulating cytokines that decreased over time. Patients in Cluster 2 and Cluster 3 had higher initial levels that decreased over time in Cluster 2 and increased in Cluster 3. Patients in Clusters 2 and 3 had higher mortality compared to Cluster 1 (Clusters 1, 2, 3: 11% vs 31% OR 3.56 (1.10–14.23), vs 54% OR 9.23 (2.89–37.22)). Cluster 3 was independently associated with in-hospital mortality (HR 5.24, p=0.005) in multivariable analysis. There were no significant differences in initial clinical severity scoring or steroid use between the clusters. Analysis of either t₁ or t₂ cytokine measurements alone or in combination did not reveal clusters with clear clinical significance.

Conclusion:

Longitudinal measurement of cytokine profiles at initiation of vasoactive medications and 24 hours later revealed three distinct cytokine signature clusters that correlated with clinical outcomes.

Introduction

Sepsis is a heterogeneous syndrome defined as the dysregulated host response to infection leading to life-threatening organ dysfunction (1, 2). Sepsis affects over 1.5 million people annually in the United States, is the most expensive condition treated in U.S. hospitals, and leads to 1 in 3 in-hospital deaths (3, 4). Identification of infection relies on clinical suspicion and blood cultures that may take days to result and are positive in as few as 17% of patients with sepsis (5). Providers rely on clinical intuition and limited severity of illness scores such as Applied Physiology and Chronic Health Evaluation-II (APACHE-II) and Sepsis-related Organ Failure Assessment (SOFA) to identify high-risk patients and predict sepsis outcomes (6, 7). Clinical trials of targeted sepsis therapies have repeatedly failed, and the mainstay of treatment remains supportive care and early antibiotics. Thus far, strategies have focused on a non-discriminant approach that lacks inclusion of biologic markers of the dysregulated host response and inadequately addresses the heterogeneity of sepsis (8–10).

Biomarkers are measurable host characteristics that reflect physiologic or pathologic processes which could be used to offer insight into aberrant host immune responses. Their use has the potential to stratify heterogeneous groups of patients into homogenous subgroups with shared clinical characteristics to predict treatment responses and improve outcomes (11–14). Modern techniques such as machine learning and cluster analysis have been successful in classifying heterogeneous diseases, such as asthma (15), acute respiratory distress syndrome (16), and interstitial lung disease (17). Numerous biomarkers have been studied in sepsis, but none have been translated into widespread biomarker-based stratification or treatment pathways (18–23).

Recent work has identified networks of multiple cytokines that correlate with disease severity, suggesting that measurement of a single biomarker at one time point vastly underrepresents the true host response (24–26). Furthermore, smaller studies have shown that dynamic changes in cytokines such as interleukin (IL) 6 rather than the absolute level at a single time point were more strongly correlated with mortality risk in patients with septic shock (24, 27–30). These studies highlight the importance of networks of biomarkers as well as temporal changes to characterize the complexity of the host response, but more work is required to combine both ideas into one dynamic model.

Developing accurate prognostication tools can lead to more targeted diagnostics and treatments in patients with septic shock. As such, we aimed to cluster patients with septic shock using 37 representative cytokines at two time points in order to identify cytokine profiles and their relationship to clinical outcomes.

Materials and Methods

Enrollment:

This prospective study was approved by the University of Chicago Institutional Review Board (IRB 18-1163: Analysis of Biomarkers and Immune Function in Hospitalized Patients with Shock) and performed in accordance with the ethical standards in the 1964 Declaration of Helsinki and its later amendments. Written consent was obtained from study subjects or their surrogates. Adult patients ≥18 years of age admitted to the Medical Intensive Care Unit at the University of Chicago who required vasoactive medication support for the treatment of shock were approached for enrollment within the first 24 hours of the diagnosis of shock. Patients with septic shock, as defined by the Sepsis-3 guidelines, were included in this study (1). Enrollment was from January 2018 to January 2020. 30 mL of whole blood was collected from each patient in sodium citrate tubes at time point t₁ (enrollment) and t₂ (24 hours after enrollment). Within 2 hours of collection, whole blood samples were fractionated by centrifugation at 2000g for 15 minutes to isolate the plasma supernatant, which was aliquoted and stored at −80°C until analysis.

Data Collection:

We performed a magnetic bead-based multiplex assay (Magnetic Luminex^® Assay from R&D Systems) that allows for simultaneous measurement of 37 cytokines: IL-1α, IL-1β, IL-1 receptor antagonist (IL-1Ra), IL-2, IL-4, IL-6, IL-7, IL-8, IL-10, IL-12p70, IL-13, IL-15, IL17E/IL-25, IL-22, IL-27, IL-33, angiopoietin, angiopoietin-2, cluster of differentiation 14 (CD14), complement component 5 (C5), chemokine C-C motif ligand 11 (CCL11)/eotaxin-1 eotaxin, eotaxin-2, chemokine C-C motif ligand 26 (CCL26)/eotaxin-3, chemokine C-C motif ligand 3/macrophage inflammatory protein 1-alpha (CCL3/MIP-1α), chemokine C-X3-C motif ligand 1 (CX3CL1), granzyme B, granulocyte colony stimulating factor (G-CSF), macrophage colony-stimulating factor (M-CSF), granulocyte-macrophage colony-stimulating factor (GM-CSF), interferon gamma (IFN-γ), lactoferrin, lipocalin-2, matrix metalloproteinase-8 (MMP-8), monocyte chemoattractant protein 1 (MCP-1), procalcitonin, serpin-E1, suppression of tumorigenicity 2 (ST2), and tumor necrosis factor alpha (TNF-α) (31).

Multiplex assays were performed on samples collected at time of trial enrollment (t₁) and 24 hours later (t₂). Time to enrollment was defined as time from septic shock development to t₁. Protein levels above or below the limit of quantification (LOQ) were assigned the value of the LOQ for analysis. The difference in cytokines between the two time points, t₂ - t₁, is reported as delta. Delta values, cytokine ratios, and percentage changes were calculated on an individual basis before statistical methods or summaries.

The electronic medical record was reviewed to extract pertinent data, including baseline demographic information, comorbid disease conditions, clinical, and mortality data. APACHE-II score was calculated on the day of ICU admission and t₂ (32). Sequential Organ Failure Assessment (SOFA) and Vasoactive-Inotrope Score (VIS) were calculated at t₁ and t₂ (7, 33, 34). Charlson Comorbidity Index (CCI) was calculated for each patient (35). Bacteremia was further classified by source of infection: line related, genitourinary, pulmonary, intraabdominal, soft tissue or bone, other, or unknown. Duration of shock was characterized using previously described clinical phenotypes: State A for patients with rapid resolution of shock within 48 hours, State B for patients with persistent shock requiring vasoactive medications greater than 48 hours, and State C for patients who did not resolve their shock and died on vasoactive medications (30).

Statistical Analysis:

Partitioning around medoids (PAM) and unsupervised random forest (RF) algorithms were used to identify clusters of patients with septic shock based on their cytokine profiles (17, 36). These analyses aim to group patients based on cytokine levels to optimize cluster homogeneity and differentiate clusters from one another. PAM cluster analysis minimizes the dissimilarity of members of each cluster and uses medoids, patients that are representative of each cluster (37). PAM was performed on raw data or random forest was used to generate a proximity matrix as input into a PAM model. In comparison to PAM alone, RF clustering (RF-derived proximity measure combined with PAM clustering) can be used to handle non-linear data with possible outliers and higher noise-signal ratio. RF clustering has been shown to be an effective method to determine the underlying structure of unlabeled data (38–40). Random forest was performed with 10000 trees without resampling or replication. A model fit plot using the silhouette width, a measure of how similar a patient is to their assigned cluster compared with neighboring clusters, determined the number of clusters that resulted in optimal fit of the data by PAM (Supplementary Table 1). PAM only and RF analysis was repeated for permutations of t₁, t₂, and delta, and their natural log transformations. For clusters with significant mortality outcomes, chi-squared values are reported and area under the curve (AUC) was calculated based on predicted probabilities of mortality using a logistic regression model. Variable importance for random forest analysis was calculated using relative Gini scoring, with higher values indicating higher importance of a variable in the model (41). Clustering analyses were performed by using the “cluster” and “randomForest” packages in R (R foundation for Statistical Computing, V4.3.0). Principal Component Analysis (PCA) was utilized to analyze correlation of principal components with cluster assignment. Calinski-Harabasz Index and Davies–Bouldin index were calculated for each cluster. Bonferroni correction was used when analyzing individual cytokine patterns.

Descriptive statistics were used for demographic and comorbidity data. We used the χ2 test or Fisher’s exact test as appropriate to compare categorical outcomes. ANOVA was used to compare continuous outcomes. Comparisons were made between clusters and between survivors versus non-survivors. Kruskal Wallis, chi-squared goodness of fit and student’s t-tests analysis were used to assess the relationship between the identified cytokine-based patient subgroups with outcomes. Survival was assessed using unadjusted log-rank testing along with multivariable Cox proportional hazards regression. Survival curves were plotted by using the Kaplan-Meier survival estimator. Cox regression analysis was performed adjusting for risk factors that showed a trend toward significance (p ≤ 0.1) in univariable analysis comparing survivors and non-survivors, or that were linked to the outcome on a biologically plausible basis. Akaike’s Information Criteria (AIC) and Bayesian Information Criteria (BIC) were calculated.

Results

Among 153 patients enrolled, 120 (78%) were classified as patients with septic shock, of whom 114 had collection of samples at both t₁ and t₂. Median age of included patients was 65.8 ± interquartile range (IQR) 18.6 years, 68% were male, and 57% were black (Table 1). Overall mortality was 31.6% (Table 2).

Table 1.

Demographic and Clinical Information by Cluster

Clinical Variable	All (N=114)	Cluster 1 (N=37)	Cluster 2 (N=42)	Cluster 3 (N=35)	P-Value
Age	65.8 ± 18.6	65.7 ± 17.7	66.3 ± 19.6	64.7 ± 15.1	0.65
BMI	27.0 ± 11.6	27.7 ± 14.0	25.6 ± 12.1	26.9 ± 8.9	0.15
Female	36 (32%)	15 (41%)	12 (29%)	9 (26%)	0.34
Race					0.12
White	37 (32%)	12 (32%)	16 (38%)	9 (26%)
Hispanic	6 (5%)	1 (3%)	4 (10%)	1 (3%)
Black	65 (57%)	23 (62%)	20 (48%)	22 (63%)
Asian	3 (3%)	1 (3%)	2 (5%)	0 (0%)
Other	3 (3%)	0 (0%)	0 (0%)	3 (9%)
Charlson Comorbidities Index	6 ± 3.8	5 ± 6.0	6 ± 3.0	6 ± 4.0	0.62
Time to enrollment (hours)	10.6 ± 10.3	11.7 ± 11.5	8.8 ± 9.8	10.9 ± 10.5	0.50
APACHE-II, at time of ICU admission	26 ± 7.8	25 ± 8.0	26 ± 8.8	27 ± 6.5	0.08
SOFA, time 1	10 ± 5.0	9 ± 5.0	11 ± 4.0	10 ± 4.5	0.08
Vasoactive-Inotrope Score, time 1	11 ± 17.0	10 ± 12.0	18 ± 25.5	10 ± 14.4	0.92

Data presented as Number (%) or Median ± Interquartile Range. ANOVA or Chi-squared used for p-values. Key: BMI - Body Mass Index, APACHE - Acute Physiology and Chronic Health Evaluation, ICU – Intensive Care Unit

Table 2.

Clinical Outcomes by Cluster

Clinical Variable	All (N=114)	Cluster 1 (N=37)	Cluster 2 (N=42)	Cluster 3 (N=35)	P-Value
Overall cytokine change	−3.6 ± 13.8%	−11.6 ± 13.3%	−21.0 ± 27.7%	21.3 ± 19.8%	<0.001
IL-10/TNF-α	9.2 ± 24.5	6.4 ± 5.8	11.0 ± 29.9	12.7 ± 22.4	0.13
APACHE-II, time 2	22.5 ± 9.0	21 ± 9.0	22 ± 8.5	25 ± 9.5	0.04
Change in APACHE-II	−3 ± 7.8	−3 ± 7.0	−2.5 ± 7.0	−3.0 ± 8.5	0.48
SOFA, time 2	8 ± 5.0	7 ± 5.0	8 ± 5.8	9 ± 4.5	0.01
Change in SOFA	−1 ± 3.0	−2 ± 4.0	−1 ± 3.0	−1 ± 4.0	0.15
Vasoactive-Inotrope Score, time 2	3.8 ± 16.4	3 ± 7.0	5.5 ± 28.5	7 ± 31.5	0.003
Change in Vasoactive-Inotrope Score	−5 ± 14.6	−7 ± 12.0	−8.5 ± 14.9	−2 ± 17.5	0.002
In-hospital mortality	36 (32%)	4 (11%)	13 (31%)	19 (54%)	<0.001
Mechanical ventilation	57 (50%)	15 (41%)	19 (45%)	23 (66%)	0.08
Renal replacement therapy	22 (19%)	8 (22%)	10 (24%)	18 (51%)	0.47
ICU length of stay	4.2 ± 8.7	4.3 ± 10.7	3.6 ± 5.5	4.7 ± 9.3	0.60

Data presented as Number (%) or Median ± Interquartile Range. Kruskal Wallis used to calculate cytokine change significance, ANOVA or Chi-squared used for all other p-values. Change in APACHE II, SOFA, and VIS is from time 1 to time 2.

Key: APACHE - Acute Physiology and Chronic Health Evaluation, ICU – Intensive Care Unit

Clustering Analysis:

Clustering was completed with t₁, t₂, t₁-t₂, delta, t₁-delta, t₁-t₂-delta data points using both PAM and random forest analysis. Delta with RF was selected for primary analysis. Of the remaining clusters, only t₂-PAM, t₁-t₂ -PAM, and t₁-t₂-delta-PAM returned a significant P-value for mortality. However, these clusters did not demonstrate clear separation and or cytokine patterns. RF on delta only provided a higher chi-squared value of 15.7 than the other significant clusters (chi-squared values 8.1–9.8) as well as a higher AUC measure (0.71 vs 0.61) (Supplementary Table 2).

Cytokine Profiles:

RF analysis on delta-only data identified three clusters (Figure 1A) with distinct cytokine patterns (Figure 1B and 1C). Between clusters, there was no significant difference in age, BMI, sex, race, or CCI (Table 1). There was no difference in t₁ SOFA, t₁ VIS or APACHE-II on ICU admission. There was no difference in receipt of steroids in the first 24 hours of study enrollment or during ICU admission. Median time from the development of septic shock to study enrollment was 10.6 ± 10.3 hours and did not differ significantly across clusters.

Figure 1:

Unsupervised random forest analysis was performed on absolute change in cytokine values over 24 hours in patients with septic shock. A distance matrix was calculated based on proximity of samples in random forest with 10,000 trees and processed with PAM, revealing three clusters (1A) with distinct cytokine patterns and clinical outcomes. 1B and 1C show violin plots of natural log transformations of average initial and change in cytokine levels by cluster with significant difference between Cluster 1 and 3 (p<0.001) and Cluster 2 and 3 (p<0.001).

Variable importance identified IL-33, IL-1α, IL-4, IFN- γ and CCL26/eotaxin-3 as the most important factors (Figure 2). Overall, Cluster 1 cytokine values fell 11.6% (IQR 13.3%, p<0.001) from t₁ to t₂ and Cluster 2 values fell 21.0% (IQR 27.7%, p<0.001). In contrast, Cluster 3 values increased by 21.3% (IQR 19.8%, p<0.001).

Figure 2:

Ranked cytokine importance for clustering determined by variable importance (mean decrease Gini coefficient)

Cluster 1 demonstrated significantly lower levels for the majority of cytokines at t₁ compared to Cluster 2 and Cluster 3 (Supplementary Table 3), with 35% of the cytokines falling on average at t₂ (Supplementary Table 4 and Supplementary Table 5). Patients in Cluster 2 and Cluster 3 had similarly high initial cytokine values at t₁ compared to patients in Cluster 1. However, Cluster 2 had significant decreases in 49% of cytokines at t₂ and no statistically significant increases, whereas Cluster 3 saw increases in 22% of cytokines and no statistically significant decreases. Compared to Cluster 1 and Cluster 3, Cluster 2 deltas were significantly more negative for most cytokines (Supplementary Table 4).

For Cluster 3, IL-2, IL-6, IL-10, IL 17E/IL-25, IL-27, G-CSF, and ST-2 decreased in value, though without statistical significance, while 21% of cytokines rose in a statistically significant manner (Supplementary Table 5). When comparing those who survived and those who did not within Cluster 3, IL-7 (p = 0.002) and IL-33 (p = 0.04) increased for survivors and IL-15 increased for non-survivors (p = 0.01). IL-6, IL-10, and G-CSF fell for both groups, but were non-significant. IL1-Ra, IL-8, IL-13, IL-17E/IL-25, IL-27, and ST-2 demonstrated a pattern where survivors’ values fell on average and non-survivors saw increased values, though these changes were not statistically significant. IL-2 fell in non-survivors but was also not significant (Supplementary Table 6).

Clinical outcomes:

There were significant differences in mortality between all three clusters. Patients in Clusters 2 and 3 demonstrated significantly higher mortality with increased odds of death compared to Cluster 1: Cluster 1 vs 2: 11% vs 31% with OR 3.56 (1.10–14.23) and Cluster 1 vs 3: 11% vs 54% with OR 9.23 (2.89–37.22). The mortality difference between Cluster 2 and 3 was significantly different as well with OR 2.60 (1.03–6.83). Log rank testing revealed significantly different survival distribution for the 3 clusters (χ² = 13, log-rank p=0.002) (Figure 3).

Figure 3:

Kaplan-Meier Curve of delta clustering by random forest analysis

In the overall cohort, there was no difference in age, BMI, gender, race, or comorbidities between survivors and non-survivors. Non-survivors had a significantly higher t₁ APACHE II (30.5 vs 24.5, p<0.0001), SOFA (12 vs 9, p<0.001), and VIS (20 vs 10, p<0.049). All three scores remained significantly higher in non-survivors at t₂. More survivors had bacteremia (45% vs 25%; p = 0.05) and a genitourinary source of infection (36% vs 8%; p = 0.002), whereas non-survivors had higher rates of pulmonary infection (26% vs 58%; p <0.001) (Supplementary Table 7). Multivariable Cox regression analysis adjusting for cluster and other explanatory variables identified in univariable analysis suggested that Cluster 3 independently predicted in-hospital mortality (HR 5.24, p = 0.005) (Supplementary Table 8).

All clusters had statistically significant decreases in VIS, APACHE-II, and SOFA score over time with the exception of Cluster 3 VIS (Table 2). When compared to Cluster 3, Cluster 1 and Cluster 2 had greater decreases in VIS: −2 compared to −7 (p = 0.005) in Cluster 1 and −8.5 (p = 0.02) in Cluster 2. Changes in SOFA and APACHE-II scoring from t₁ to t₂ were not different between individual clusters, though t₂ SOFA (Cluster 1: 7, Cluster 2: 8, Cluster 3: 9; p = 0.01) and t₂ APACHE-II scores (Cluster 1: 21, Cluster 2: 22, Cluster 3: 25; p = 0.04) were statistically different. Compared to Cluster 3, Cluster 1 had significantly lower t₂ VIS (7 vs 3, p = 0.09) and t₂ SOFA scoring (9 vs 7, p = 0.04).

There was a significantly higher rate of invasive mechanical ventilation (MV) for patients in Cluster 3 compared to Cluster 1: 66% vs 41% (p = 0.03) with OR 2.76 (1.06–7.44). Clusters also demonstrated a significantly different proportion of shock states (p = 0.047), with Cluster 3 demonstrating a higher percentage of State C shock patients (Supplementary Table 9), those who never recovered and died while on vasoactive medications (37%, compared to 8% in Cluster 1 and 21% in Cluster 2) (30). There was no significant difference in intensive care unit length of stay (ICU-LOS) or new renal replacement therapy (RRT). There was a significant difference in bacteremia among the clusters with Cluster 2 demonstrating the highest rate and Cluster 3 demonstrating the lowest (55% and 23%, p = 0.01). There was no difference in anatomic site of infection (Table 3).

Table 3.

Bacteremia Status and Source of Infection by Cluster

	All (N=114)	Cluster 1 (N=37)	Cluster 2 (N=42)	Cluster 3 (N=35)	P-Value
Bacteremia	44 (39%)	13 (35%)	23 (55%)	8 (23%)	0.01
Source of Infection					0.65
Line Related	5 (4%)	2 (5%)	2 (5%)	1 (3%)
Genitourinary	31 (27%)	14 (38%)	11 (26%)	6 (17%)
Pulmonary	41 (36%)	13 (35%)	13 (31%)	15 (43%)
Intra-abdominal	18 (16%)	2 (5%)	10 (24%)	6 (17%)
Soft Tissue or Bone	11 (10%)	3 (8%)	4 (10%)	4 (11%)
Other	2 (2%)	1 (3%)	0 (0%)	1 (3%)
Unknown	6 (5%)	2 (5%)	2 (5%)	2 (6%)

Data presented as Number (%). ANOVA used for p-values.

Sensitivity Analysis:

Natural log transformation and scaled transformation of cytokine data yielded nearly identical clustering results. Clustering with the top N variables identified by variable importance yielded similar cluster results: top 10 cytokines yielded 83% similarity in cluster assignment, which increased to 90% with the top 20 cytokines. Top five principal components in PCA did not demonstrate strong correlation with cluster assignment (Supplementary Table 10).

Discussion

Cluster analysis of the change in cytokine values revealed novel clusters of patients with septic shock. We identified three cytokine signature clusters that were associated with significant differences in mortality, mechanical ventilation, and shock phenotype, and that were independent of initial severity of illness.

Patients in Cluster 1 had the lowest average initial levels of circulating cytokines (t₁) which further decreased over time, and had the lowest mortality. Patients in Cluster 2 and Cluster 3 had higher initial levels (t₁) and higher mortality. However, Cluster 3 cytokines further increased over time (delta Cluster 3 > delta Cluster 2) and were associated with the highest mortality. The finding that higher cytokine values are associated with higher mortality has been described previously (42–44). However, clustering on delta values offered different subgroupings than static thresholds alone, suggesting that clustering based on single time points may exclude important phenotypic information and does not achieve the full prognostic value of cytokines. Thus, it can be theorized that a patient’s changes in cytokines are an indirect measure of evolving immune regulation or dysregulation and reflect either clinical recovery or deterioration. More studies are needed to elucidate specific timing and clinical utility.

Beyond evaluating the overall pattern of cytokine changes over time, the identities of the cytokines themselves may provide insight into functional and dysregulated host responses. Within Cluster 3, IL-7 levels increased in survivors (p = 0.002), while a rise in IL-15 levels was associated with mortality (p = 0.01). Previous data has demonstrated that IL-7 promotes lymphocyte expansion and enhanced function leading to improved outcomes (28) while IL-15 may be triggering overactivation of natural killers and cytotoxic CD8 T cells (45). In addition, while just under 40% of Cluster 3 were State C Shock (died on vasoactive medications), a similar percentage were State A Shock (resolved < 48 hours). IL1-Ra, IL-8, IL-13, IL-17E/IL-25, IL-27, and ST-2 demonstrated a pattern where Cluster 3 survivors’ values fell on average and non-survivors saw increased values, though these changes were not statistically significant. These data suggest that a rapid increase followed by a rapid decrease in these specific cytokines may be particularly important to a regulated host response, a hypothesis that should be tested in future work. Furthermore, the most important cytokines in defining the clusters were a mix of inflammatory cytokines and mediators (IL-1α, granzyme B, IFN- γ) as well as cytokines associated with type 2/TH2 responses (IL-4, IL-33, and CCL26/eotaxin-3) (Figure 2). Increases in IL-33 in Cluster 3 patients were also statistically associated with survival (p=0.04). Taken together, these data highlight the potential importance of activation of type 2/Th2 responses to counterregulate proinflammatory sepsis responses as has been previously proposed (46–49).

Our results highlight the importance of dynamic measurements of the host immune response in sepsis and could inform future studies to define dynamic biomarkers of prognostic significance or sepsis phenotypes for customized therapies. Biomarkers at single time points may not accurately predict a patient’s clinical course, as was seen in our inability to identify reproducible clusters using single measurements of biomarkers. Serial assessments may identify patients with a resolving inflammatory response who might be harmed by immunosuppressive therapies and/or benefit from immune stimulation (50). On the other hand, identifying patients with an inflammatory immune profile, such as in Cluster 3, may guide targeted clinical trials with readily available anti-cytokine biologics for patients at highest risk of death. The ability to identify both low-risk and high-risk cytokine signature clusters and intervene effectively has the potential to transform care of patients with septic shock.

Our study has several limitations. Patients were predominantly male, Black, and enrolled at a single tertiary referral center, thus limiting generalizability. The study, while prospective, was observational in nature, and did not incorporate additional intervention beyond standard of care. The uncertainty in cluster size and estimation of variance in the absence of resampling remain a concern and will require further validation studies. While our unsupervised approach found differences in cytokine values, the physiologic relevance of micro- and picomolar changes are not fully understood though they are hypothesis generating. Finally, rapid measurement of a slate of 37 cytokines is not feasible for routine clinical care, but future work may identify specific cytokines of significance. A limited panel of these important markers may be amenable to rapid quantification using bedside tools which can then be used in future prospective studies (30).

Conclusions

Cluster analysis on the differences in cytokine levels between two time points revealed novel cytokine signature clusters with distinct clinical outcomes in patients with septic shock. Identification of such clusters may offer the potential of precision medicine in sepsis and septic shock. Further studies are needed to validate cytokine profiles and interventions.

Key Points:

Question:

Can dynamic cytokine networks identify homogenous subgroups of patients with septic shock?

Findings:

Random forest cluster analysis of the difference in cytokine levels over two time points revealed novel clusters with distinct clinical outcomes, independent of clinical severity scoring. A pattern of persistently elevated cytokines over time was associated with higher mortality.

Meanings:

Clustering based on dynamic changes in cytokines may reveal novel approaches for targeted immune modulation to improve sepsis outcomes.