Biomarker-Based Clustering to Identify Distinct Host Response Signatures in Hospitalized Patients With Suspected Infection
Document Type
Conference Proceeding
Publication Date
5-2025
Publication Title
American Journal of Respiratory and Critical Care Medicine
Abstract
RATIONALE: Hospitalized patients with suspected infection experience highly heterogeneous complications and outcomes, including sepsis and death. Differences in host responses across these patients are poorly understood. Thus, we sought to determine the breadth of biomarker signatures and identify distinct molecular phenotypes to better elucidate this population’s heterogeneity. METHODS: Adults with an Emergency Department blood culture order who were later admitted to one of four medical centers between August 2018 and July 2022 were eligible for this retrospective analysis of prospectively collected data. Biomarkers were measured from plasma samples obtained during routine care early in the stay and close to blood culture order time. For phenotype derivation, we first performed principal component analysis (PCA) on the 29 protein biomarkers to eliminate multicollinearity. We further reduced the feature dimensions using uniform manifold approximation and projections (UMAP) and identified phenotypes in a 3-dimensional space with K-means clustering. Lastly, outcomes (primary outcome: discharge on or after day seven or death), comorbidities, and organisms were compared across phenotypes using descriptive statistics and logistic regression. RESULTS: The 3,802 patients included formed six clusters uniquely characterized by multivariable biomarker distributions (Figure). Cluster 1 contained the most patients (n=1,159) and experienced the best outcomes (25.9%). Additionally, in Cluster 1, the average values of most biomarkers were less than the population mean (Figure). A similar trend of averages less than the population mean characterized Cluster 2 (n=670) but with notably higher values of C-reactive protein (CRP) (Figure). Cluster 3 (n=558) experienced the worst outcomes (52.7%) and was driven by high tissue factor and sTREM-1 with proinflammatory cytokines near the population mean (Figure). Contrastingly, Cluster 4 (n=337) had lower tissue factor and sTREM-1 values but high GM-CSF and IP10 values (Figure). Logistic regression to determine independent clinical factors associated with cluster membership showed Cluster 4 was driven by viral organisms (OR: 4.3 [1.8,9.9]). Cluster 5 (n=663) and 6 (n=415) membership was driven by gram-positive and gram-negative organisms, respectively, (OR: 1.7 [1.3, 2.2] & OR: 3.1 [2.4, 3.9]), with Cluster 5 experiencing worse outcomes (51.4% vs 40.7%) despite lower levels of proinflammatory cytokines (Figure). CONCLUSIONS: Plasma protein biomarkers measured early in the stay revealed distinct host response signatures with distinct outcomes. Importantly, the phenotype signatures were not distinguishable using standard clinical features. Our schema may offer a novel, biologically-informed, approach to classifying hospitalized patients that may aid identification of novel therapies.
Volume
211
First Page
A2963
Last Page
A2963
Recommended Citation
Spicer A, Bhavani SV, Lopez-Espina C, Watson GL, Bhargava A, Schmalz L, et al. [Sims MD, Demarco C, Halalau A, Maddens N, Espinosa A, Davila F, Davila H, Smith S]. Biomarker-based clustering to identify distinct host response signatures in hospitalized patients with suspected infection. Am J Respir Crit Care Med. 2025 May;211:A2963. doi:10.1164/ajrccm.2025.211.Abstracts.A2963
DOI
10.1164/ajrccm.2025.211.Abstracts.A2963
Comments
American Thoracic Society (ATS) International Conference, May 16-21, 2025, San Francisco, CA