Integrated immune and endothelial profiling predicts 90-day mortality in postoperative sepsis and septic shock

Summary

Background

Sepsis and septic shock remain major causes of mortality in critically ill postoperative patients, largely because of the lack of reliable biomarkers for early risk stratification. The interplay between immune dysfunction and endothelial activation is key in the progression to multiorgan failure, however phenotypic characterisation of circulating endothelial subpopulations remains limited.

Methods

A Prospective multicentre study included 219 postoperative patients (Non-septic ICU patients, sepsis, septic shock). Peripheral Blood Mononuclear Cells were analysed using high-dimensional spectral flow cytometry. Both supervised gating strategies and high-dimensional unsupervised analyses (UMAP and FlowSOM) were applied to identify immune and endothelial cell subsets. Associations with 90-day mortality were assessed using univariate and multivariate Cox proportional hazards models, refined with LASSO-Cox regression, and integrated into a risk score. The predictive performance of this cellular risk score was compared with SOFA and APACHE II scores using ROC curves and survival analysis. Findings were further validated using publicly available single-cell RNA datasets.

Findings

Two B cell subsets (plasmablasts/IgG⁺ and IgD⁻/IgM⁺ memory B cells) and endothelial cluster (CD36⁻/CD16⁺) were independently associated with 90-day mortality. The integrated risk score stratified patients into three groups with significantly different survival outcomes (log-rank p = 0.0009), outperformed SOFA and APACHE II (AUC 0.935 vs. 0.751–0.804). Transcriptomic validation confirmed the presence and dysfunction of endothelial clusters, reinforcing their prognostic relevance.

Interpretation

The combination of immune and endothelial profiling provides a robust cellular signature that improves the prognostic stratification in postoperative sepsis. These biomarkers may support treatment and guide therapeutic strategies aimed at restoring immune-endothelial homoeostasis.

Funding

This work was supported by the Instituto de Salud Carlos III [grant numbers: PI24/00754, FI25/00242 and CIBERINFEC CB21/13/00051], Junta de Castilla y León [GRS 2782/A2/2023, GRS 2804/A1/2023].

Research in context.

Evidence before this study

We searched PubMed and relevant literature databases for studies published up to the time of study conception investigating immune and endothelial biomarkers in sepsis and septic shock. Search terms included combinations of “sepsis”, “endothelial dysfunction”, “circulating endothelial cells”, “immune profiling”, “flow cytometry” and “single-cell analysis”. We considered original research articles, reviews and meta-analyses without language restrictions.

Existing evidence indicates that sepsis is associated with profound immune dysregulation and endothelial injury, both of which contribute to disease progression and mortality. Previous studies have reported associations between circulating endothelial cells, endothelial progenitor cells, and soluble endothelial activation markers (such as angiopoietins, ICAM-1, VCAM-1 and thrombomodulin) with disease severity and outcomes. Similarly, immune alterations including changes in lymphocyte subsets and plasmablast expansion have been described in critically ill patients. However, most studies have assessed immune and endothelial compartments separately, often focussing on isolated biomarkers with heterogeneous and sometimes conflicting results. Integrated high-dimensional profiling approaches combining immune and endothelial phenotyping in a unified framework remain limited, and no consensus cellular signature for risk stratification in sepsis has been established.

Added value of this study

This study provides a comprehensive integrated immuno-endothelial profiling of critically ill patients with sepsis and septic shock using high-dimensional flow cytometry combined with unsupervised clustering approaches. We identified distinct immune and endothelial cell subsets associated with 90-day mortality, including two B-cell populations (plasmablasts/IgG⁺ memory and IgD⁻/IgM⁺ memory) and a specific dysfunctional endothelial cluster (CD36⁻/CD16⁺). Importantly, we developed a composite risk score based on these cellular subsets, which enabled robust stratification of patients into distinct prognostic groups and demonstrated higher predictive performance compared with established clinical severity scores (SOFA and APACHE II).

Implications of all the available evidence

Taken together with existing literature, our findings support the concept that sepsis is driven by a coordinated immune-endothelial dysregulation rather than isolated alterations in individual biomarkers. The identification of a reproducible immune-endothelial signature suggests that cellular-level profiling may provide clinically relevant prognostic information beyond conventional scoring systems. These results have potential implications for precision medicine in sepsis, as they may improve early risk stratification, guide patient monitoring and support selection of high-risk patients for targeted interventions or clinical trials. Furthermore, the identified endothelial and immune cell subsets represent potential mechanistic targets for future therapeutic strategies aimed at restoring immune-endothelial homoeostasis.

Introduction

Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection and represents one of the leading causes of admission and mortality in intensive care units (ICU) worldwide.¹ In Europe, the incidence of sepsis is estimated to be more than 3.4 million cases per year; 700,000 of these patients do not survive hospitalisation and one third of the survivors die during the first year after hospitalisation.^2,3 In the absence of early intervention, the disease can swiftly advance to septic shock (SS) or acute multiple organ failure, defined by severe haemodynamic instability and is associated with substantial mortality.^4,5 Despite this high burden, early diagnosis and risk stratification remain limited, as current clinical scores and laboratory markers lack the sensitivity and specificity to predict outcomes.

Among the most relevant pathological changes, sepsis is characterised by a dysregulated immune response and widespread vascular damage, where the endothelium and immune cells form a key inflammatory axis that drives progression to multiorgan failure.⁶ Both cell types act not only as targets of inflammation but also as amplifiers of the host response.⁷ Their crosstalk induces an activated endothelial phenotype with proadhesive, proinflammatory and prothrombotic features.^8,9 Endothelial cells (ECs) function as both immune sensors and effectors, driving leucocyte adhesion, inflammation and coagulation. At the same time, the immune system shifts from an initial hyperinflammatory state to subsequent immunosuppression, increasing susceptibility to secondary infections and mortality.¹⁰ Thus, immuno-endothelial profiling may help identify prognostic cellular biomarkers and improve risk stratification in sepsis.

Given the diagnostic and therapeutic challenges posed by sepsis and the limitations of traditional severity indices for early stratification of critically ill patients,¹¹ there is growing interest in identifying cellular biomarkers that better reflect the patient's immune and endothelial status.¹² Building on this, previous studies have explored endothelial dysfunction in sepsis using both circulating cells and soluble biomarkers. Circulating endothelial cells (CECs) and endothelial progenitor cells (EPCs) have been associated with disease severity and mortality, reflecting ongoing vascular injury and repair processes.¹³ In parallel, soluble markers such as angiopoietin-2, soluble thrombomodulin and adhesion molecules ICAM-1 and VCAM-1, have been widely investigated as indicators of endothelial activation and barrier disruption.¹⁴ More recently, cell-surface markers including CD146, related to endothelial integrity and activation,¹⁵ and CD36, involved in inflammation and microvascular dysfunction, have also been implicated in vascular alterations in inflammatory conditions.¹⁶ However, these approaches have largely focused on isolated biomarkers, often yielding heterogeneous results and providing a limited and fragmented view of endothelial involvement in sepsis.

Although most studies have analysed these components separately, integrated immuno-endothelial profiling in sepsis remains largely unexplored. In this context, high-dimensional flow cytometry has emerged as a powerful tool for simultaneously characterising immune system cells and circulating ECs.^17,18 This technology enables multiparametric single-cell analysis and helps to identify cell subpopulations linked to disease progression and prognosis. The joint assessment of the immune system and ECs in the peripheral blood provides an integrative view of host response and severity progression, constituting an innovative framework to improve risk stratification in critically ill patients.¹⁹

Therefore, this dual immuno-endothelial approach aims to identify and quantify immune and ECs subpopulations in the peripheral blood of critically ill patients with sepsis and SS using high-dimensional flow cytometry, determine their independent association with 90-day mortality develop and validate a composite risk score based on these cellular subsets, and compare its prognostic performance with conventional severity indices (SOFA, APACHE II).

Methods

Ethics

This study was conducted in accordance with the fundamental principles established in the Declaration of Helsinki, the Convention of the European Council related to human rights and biomedicine, the Ethical Guidelines for Health-related Research Involving Humans by the Council for International Organisation of Medical Sciences of the World Health Organisation (WHO), and the requirements established by the Spanish legislation for biomedical research, the protection of personal data and bioethics. Ethics approval was obtained from the Scientific University Clinical Hospital of Valladolid (Spain) (PI-18-972), the Ethics Committee for Clinical Research of University Clinical Hospital of Toledo (Spain) (CEIC-466-2019) and University Clinical Hospital of Bellvitge (Spain) (PR295/22). Finally, the development and validation of the prognostic risk score adhered to the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) guidelines.²⁰

Study design and population

This prospective, observational, analytical and multicentre study was conducted in two independent cohorts of: i) a discovery cohort of 219 adult patients undergoing surgery from the intensive care units (ICUs) of University Clinical Hospital of Valladolid (Spain), n = 131, University Clinical Hospital of Toledo (Spain), n = 33 and University Clinical Hospital of Bellvitge (Spain), n = 55, recruited between 01/07/2021 and 23/01/2025 and ii) an external cohort of patients using the scRNA-seq dataset,²¹ downloaded from the Broad Institute Single Cell Portal/Single Cell EMBL-GEO repository (Broad Institute Single Cell Portal, SCP548), consisting of 29 patients with sepsis and 36 controls.

Non-septic ICU patients included postoperative patients who did not develop sepsis or shock. This group was chosen to match the surgical and critical care exposure of the other groups, while excluding patients with infection-related organ dysfunction. Sepsis and SS were diagnosed according to the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3 definition).²² All cases were reviewed independently by two physicians before patients were assigned to their respective groups. Written informed consent was obtained from each patient or their legal representative before recruitment. The study adhered to the ethical principles described in the World Medical Association Declaration of Helsinki. In addition, this study was conducted and reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement for observational cohort studies. Given the focus on biomarker assessment, we also followed the STROBE-ME extension for molecular epidemiology studies.²³ Fig. 1A shows the STROBE flowchart summarising the study characteristics, while Fig. 1B illustrates the study design.

Fig. 1.

Study flowchart and design. A shows STROBE flowchart summarising the study characteristics. A total of 219 critically ill patients were included, comprising 77 Non-septic ICU patients, 62 sepsis patients and 80 patients with SS. The cohort was divided into a training set (n = 153) for model development and a test set (n = 66) for validation. B, study design. Peripheral blood mononuclear cells (PBMCs) were isolated from patients classified into three clinical groups (non-septic ICU patients, sepsis and SS) for subsequent antibody staining and analysis using spectral cytometry (Cytek Aurora). Data were processed with unsupervised analysis algorithms and multivariate statistical methods, enabling the identification of cellular subsets associated with 90-day mortality and their integration into predictive risk stratification models.

Clinical data and sample collection

Epidemiological and clinical data were extracted from the patient's medical records. Peripheral blood samples were collected in EDTA tubes and were sent directly to the HCUV Research Support Unit for processing. Peripheral Blood Mononuclear Cells (PBMCs) were isolated from peripheral venous blood by Ficoll-Hypaque density gradient centrifugation (IBIAN Technologies; Cat number: P04-60500). The cells were immediately cryopreserved in freezing medium (10% DMSO (Sigma-Adrich; Cat number: D2650-100 ML) and 90% FBS (Biowest; Cat number: S1810) and stored in liquid nitrogen (−196 °C) until used for further staining.

Blood sampling timing:

• -Non-septic ICU patients: Blood samples were obtained within the first 24 h after surgery.
• -Sepsis/septic shock patients: Blood sampling was performed within the first 24 h following the clinical diagnosis of sepsis or SS.

Antibody staining and cell acquisition

After isolation, PBMCs were stained using a customised high-dimensional flow cytometry panel based on OMIP-069,²⁴ originally designed for the characterisation of immune subsets populations. To extend its applicability to endothelial cell analysis and future combined analyses with immune and endothelial cells, the panel was modified by removing certain functional immune markers and incorporating 13 additional antibodies directed against endothelial-associated molecules. Although the final panel comprised 40 antibodies in total (27 immune-related and 13 endothelial-related), only the endothelial-related markers were considered for the present study (Supplementary Table S1). Cell acquisition was performed using a full-spectrum spectral flow cytometer (Cytek Aurora, 5 lasers, 67 detectors) at the Institute of Biomedicine and Molecular Genetics (IBGM, University of Valladolid/CSIC). Spectral flow cytometry captures the full emission spectra of each fluorochrome, enabling simultaneous discrimination of fluorochromes that are difficult to resolve with conventional cytometry.

The spectral compatibility of the fluorochrome combination was assessed using the similarity and complexity indices provided by SpectroFlo® software (Cytek Biosciences, v3.3.0). Fluorochromes with similarity indices ≤0.98 were considered sufficiently distinct to ensure optimal spectral separation. All antibodies were titrated through serial dilutions, and their staining performance was evaluated using positive/negative population histograms and staining index calculations. Spectral unmixing was conducted with SpectroFlo® to correct for cellular autofluorescence and improve signal accuracy. The accuracy of the unmixing matrix was verified by analysing NxN plots and confirming appropriate marker resolution through monochromatic and multicolour histograms in OMIQ® Data Science software (Omiq, Inc., 2022).

Prior to antibody labelling, dead cells were excluded using the Live/Dead Fixable Blue Dead Cell Stain Kit (Molecular Probes, Thermo Fisher Scientific). Brilliant Stain Buffer and True-Stain Monocyte Blocker were added to minimise nonspecific binding and optimise fluorescence intensity. Platelets were excluded from the analysis using Syto 13 (Molecular Probes, Thermo Fisher Scientific; Cat number: S7575), a DNA-permeable green fluorescent dye. Intracellular staining for von Willebrand factor (VWF) was performed using the FIX & PERM™ Cell Permeabilization Kit (Thermo Fisher Scientific; Cat number: GAS004) following the manufacturer's instructions. Cells were subsequently fixed in 1% paraformaldehyde prepared in FACS buffer (500 mL PBS, 10 mL filtered FBS, 0.1 g NaN₃, and 2.5 mL sterile EDTA) for 10 min in the dark, then washed and stored at 4 °C until acquisition within 24 h. Because the staining included an intracellular marker, the protocol was re-optimised, and all necessary quality controls, including FMO, isotype and viability controls, were repeated to ensure specific detection of the target endothelial populations under the modified panel conditions.

Cytometry data

Spectral cytometry data were analysed using hierarchical gating strategies in OMIQ® following two complementary analytical strategies: (i) a supervised, hypothesis-driven approach based on conventional hierarchical manual gating, and (ii) an unsupervised, data-driven clustering approach using FlowSOM. Antibody staining and cell acquisition detailed in the Supplementary Material.

For the supervised analysis (manual gating), data were normalised and subjected to quality control using the PeacoQC algorithm to automatically remove events affected by technical artifacts such as clogs or signal fluctuations. Subsequent manual filtering excluded debris and doublets, retaining only viable CD31⁺/CD45^- ECs and CD45⁺ immune cells. To further validate the endothelial identity of the identified population, CD45⁻/CD31⁺ cells were flow-sorted and subjected to morphological assessment and short-term culture. Sorted cells displayed size and morphology consistent with circulating endothelial cells and maintained cellular integrity after culture, suggesting that the analysed events correspond to viable endothelial cells rather than debris or microparticles. Due to software limitations in handling large datasets, random subsampling was applied to obtain a total of 5 million events, ensuring balanced representation of all study cohorts. Data were then scaled and transformed (log₁₀) to normalise cell proportions across samples before downstream analyses. For the unsupervised analysis, dimensionality reduction was performed using Uniform Manifold Approximation and Projection (UMAP)²⁵ to visualise high-dimensional structure, followed by clustering with the FlowSOM²⁶ (unsupervised analysis) algorithm to identify distinct ECs subsets, including both expected and previously uncharacterised populations. FlowSOM clusters were subsequently grouped into metaclusters using the elbow method. The resulting clusters were projected onto the UMAP embedding to assess spatial coherence. Cluster robustness was evaluated through bootstrap resampling to confirm stability and reproducibility.

UMAP was executed with the following parameters: number of neighbours = 15, minimum distance = 0.4, metric = Euclidean, components = 2, random seed = 8287, learning rate = 1, number of epochs = 200, and spectral initialisation. FlowSOM clustering used the same marker set as UMAP, with 10 training iterations, Euclidean distance, the elbow method for metaclustering and a random seed of 190. The resulting FlowSOM clusters were projected onto the UMAP embedding to confirm spatial correspondence among metaclusters.

Statistical analysis

Classical statistical analyses were performed using R software (version 4.4.1). Categorical variables were compared using χ² or Fisher's exact tests, while continuous variables that were not normally distributed, as assessed by the Shapiro-Wilk test, were analysed using the Mann-Whitney U tests. For comparisons across more than two groups, the Kruskal–Wallis test with Benjamini-Hochberg correction was applied. Statistical p-value <0.05 was considered statistically significant. To evaluate predictive performance, the cohort was randomly divided into training (70%) and test (30%) sets, with 90/10 cross-validations performed within the training cohort to optimise model parameters. Sample size was calculated to detect a small-to-medium effect (Cohen's f = 0.20, equivalent to Cohen's d ≈ 0.4 for two-group comparisons) among three independent groups, with a significance level of 0.05 and 90% power. The estimated required sample size was 207 patients. Ultimately, 219 patients were enrolled across the three groups (Non-septic ICU patients: n = 77; Sepsis: n = 62; SS: n = 80). Pregnant women, terminally ill patients, patients with limited therapeutic effort, and patients who met clinical criteria for SS but had negative microbiological cultures were excluded.

To evaluate the association between immune and endothelial cell subsets and 90-day mortality, univariate Cox proportional hazards (CoxPH) regression was performed for all clusters. Significant populations were further refined using LASSO-Cox regression to minimise overfitting, followed by recursive feature selection based on Akaike Information Criterion (AIC) to define the optimal multivariate model. The final model was adjusted for age and sex and included three key cellular populations. A weighted risk score was constructed by applying the Cox β-coefficients to each patient's log-transformed cell proportions. Patients were stratified into low-, intermediate- and high-risk groups. Survival outcomes were assessed using Kaplan-Meier curves and the log-rank test, while mortality distributions across risk groups were compared with Fisher's exact test.

Model discrimination was quantified using Receiver Operating Characteristic (ROC) curves generated with the pROC package. Area under the curve (AUC) values were classified as fair (>0.7), good (>0.8) and excellent (>0.9), with 95% confidence intervals calculated. Confusion matrices and derived metrics, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value and balanced accuracy were computed to evaluate classification performance.

scRNA-seq external validation

For external validation, single-cell RNA-seq datasets were analysed using Seurat and differential expression analyses were performed using the FindMarkers function. Marker enrichment and pathway analyses were conducted using the KEGG, Gene Ontology (GO) and Reactome databases, implemented through the clusterProfiler, org.Hs.eg.db, ReactomePA and enrichplot packages. The results were visualised using volcano plots to significantly upregulated or downregulated genes within the endothelial and immune clusters.

Role of the funders

There was no study-specific funding source for this work. Individual authors received salary or fellowship support from their respective institutions and funding bodies; however, these funders had no role in the study design, data collection, data analysis, data interpretation or writing of the report. All authors had full access to the data and take responsibility for the integrity and accuracy of the analysis.

Results

Patient characteristics

Baseline characteristics of the 219 patients are summarised in Table 1. Compared to survivors, non-survivors were older (median 72 vs. 78 years, p = 0.006) and had a higher prevalence of renal disease (4.8% vs. 18.8%, p = 0.004). At diagnosis, they showed lower lymphocyte counts (0.84 vs. 0.60 × 10⁹/L, p = 0.003), higher CRP levels (144 vs. 283 mg/L, p = 0.024), and elevated creatinine (0.98 vs. 1.10 mg/dL, p = 0.021). Total bilirubin was slightly lower in non-survivors (0.69 vs. 0.75 mg/dL, p = 0.018). Disease severity was reflected by higher SOFA (1 vs. 5, p < 0.001) and APACHE II (11 vs. 17, p < 0.001) scores. Clinical outcomes, with longer ICU stay (3 vs. 7 days, p = 0.002), longer hospital stay (15 vs. 14 days, p = 0.005), higher need for intubation (22% vs. 5%, p = 0.021) and markedly increased mortality at both 28 (0% vs. 59%, p < 0.001) and 60 days (0% vs. 94%, p < 0.001).

Table 1.

Descriptive table.

	Alive (N = 187)	Dead (N = 32)	p-value
Characteristics
Category	56 (29.8%)	56 (29.8%)	<0.001
Male, n (%)	113 (60.4%)	21 (65.6%)	0.554
Female, n (%)	74 (39.6%)	11 (34.4%)	0.554
Age, median (IQR)	72 (16)	78 (11)	0.006
Comorbidities
Smoker	32.0 (17.0%)	5.0 (15.6%)	0.845
Cardiovascular disease	65.0 (34.6%)	15.0 (46.9%)	0.181
Diabetes mellitus	54.0 (28.7%)	8.0 (25.0%)	0.665
Neurological disease	27.0 (14.4%)	2.0 (6.2%)	0.210
Arterial hypertension	104.0 (55.3%)	20.0 (62.5%)	0.449
Hepatopaties	8.0 (4.3%)	1.0 (3.1%)	0.765
Pulmonary disease	21.0 (11.2%)	8.0 (25.0%)	0.095
Renal disease	9.0 (4.8%)	6.0 (18.8%)	0.004
Obesity	41.0 (21.8%)	2.0 (6.2%)	0.040
Cancer last 3 months	41.0 (21.8%)	11.0 (34.4%)	0.122
Biochemistry at diagnosis
Glucose (mg/dl)	148.5 (52.7)	158.5 (57.0)	0.458
Leukocytes (103 x cells/mm3)	10.850 (7.700)	11.385 (11.957)	0.624
Lymphocytes (103 x cells/mm3)	0.840 (0.665)	0.600 (0.355)	0.003
Neutrophils (103 x cells/mm3)	9.24 (6.90)	10.21 (8.79)	0.751
CRP (mg/l)	144.39 (216.73)	283.00 (222.30)	0.024
Lactate (mm)	1.44 (0.87)	1.86 (1.37)	0.009
Haematocrit (%)	31.4 (7.2)	30.3 (7.7)	0.372
Sodium (mEq/l)	137.00 (5.00)	136.00 (5.50)	0.189
Potassium (mEq/l)	3.9 (0.7)	3.8 (0.5)	0.629
Creatinine (mg/dl)	0.98 (0.73)	1.10 (1.32)	0.087
Total bilirubin (mg/dl)	0.69 (0.67)	0.75 (0.61)	0.998
Platelets (103 x cells/mm3)	188.0 (111.7)	290.5 (252.7)	0.051
SOFA Score	1 (4)	5 (4)	<0.001
Apache II Score	11 (8)	17 (9)	<0.001
Time course and outcome
Length of ICU stay, median (IQR)	3 (4)	7 (11)	0.002
Length of hospital stay, median (IQR)	15 (19)	14 (25)	0.722
Intubation time	0 (1)	1 (2)	0.021
Mortality [% at 28 days]	0 (0.0%)	19 (59.4%)	<0.001
Mortality [% at 60 days]	0 (0.0%)	30 (93.8%)	<0.001

Non-survivors were older, had more renal disease, lower lymphocyte counts, higher CRP and creatinine, and higher SOFA and APACHE II scores. Significant differences (p < 0.05) were observed for category, age, renal disease, obesity, lymphocytes, CRP, lactate, SOFA and APACHE II scores and clinical outcomes including hospital and ICU length of stay and 28- and 60-day mortality.

Continuous variables are represented as median and interquartile range (IQR); categorical variables are represented as number (n) and percentages (%). p-values in bold indicate statistical significance. INR: International Normalised Ratio; ICU: Intensive Care Unit.

Supervised gating strategy and clustering for the identification of immune subsets

High-dimensional spectral flow cytometry was applied to a cohort of 219 critically ill patients to characterise the peripheral immune and ECs populations and identify subsets associated with 90-day mortality. As shown in Supplementary Figure S1, ECs events were normalised as the ratio of endothelial cells to total acquired cells to account for inter-patient variability in total cell counts, showing differences across groups sepsis and SS, with comparisons between Non-septic ICU patients, sepsis and SS groups indicated (Kruskal–Wallis p = 0.16).

Supplementary Figure S2 shows immune populations defined by a classical supervised gating strategy based on sequential selection using established surface markers, including T cells (CD4⁺, CD8⁺, γδ T, Tregs), B cells (naive, memory, plasmablasts), NK cells, NKT cells and monocyte subsets (classical, non-classical, intermediate). As shown in Supplementary Figure S3, immune composition varied progressively with clinical severity. Non-septic ICU patients displayed balanced T and memory B populations, whereas sepsis and SS showed marked plasmablast expansion and enrichment of IgD⁻ B-cell subsets (plasmablasts/IgG⁺ and memory IgD⁻/IgM⁺), both associated with mortality.

High-dimensional analysis and clustering of circulating endothelial cells

To characterise the cellular landscape, dimensionality reduction using UMAP was first performed on the combined CD45⁺ immune and CD45⁻/CD31⁺ endothelial cell compartment, revealing 12 biologically coherent clusters corresponding to canonical immune and endothelial populations (Fig. 2A). Subsequently, the endothelial compartment was isolated and analysed in greater detail using UMAP and FlowSOM clustering, identifying 13 distinct endothelial clusters (Fig. 2B), each defined by a unique marker expression profile (Fig. 2C). The heatmap illustrates the differential expression of endothelial markers across the identified clusters. Cluster 1 showed high PV1, along with CD36, HLA-DR and CD105, and low CD16. Clusters 2 and 3 were mainly defined by CD36, with cluster 2 also expressed PV1 and low CD16. Clusters 4 and 7 shared high Ephrin and CD36, while Cluster 4 had a lower VWF. Cluster 5 exhibited strong HLA-DR together with CD36 and CD105, whereas cluster 6 was predominantly HLA-DR. Cluster 8 expressed CD36 and VWF, while cluster 9 showed HLA-DR with low CD16 and PDPN. Endothelial Cluster 10 was defined by CD16 and lower CD36, and cluster 11 by CD32B. Cluster 12 co-expressed HLA-DR, CD36, CD105 and PV1, with reduced levels of CD16 and PDPN. Finally, Cluster 13 displayed high VWF, CD309 and CD16, contrasting with lower PV1 and CD36. Data structure visualisation was achieved through dimensionality reduction using UMAP, facilitating the representation of cluster distribution in the phenotypic space. Fig. 2D shows a heat map of endothelial clusters in different clinical groups (Non-septic ICU patients, sepsis and SS), showing variation in cluster distribution according to disease severity. The CD36⁻/CD16⁺ cluster was increased in patients with SS, in contrast to clusters 2 and 3, which are more present in Non-septic ICU patients. In Supplementary Figure S4 and Supplementary Figure S5, the UMAP of each endothelial marker can be seen.

Fig. 2.

High-dimensional analysis and clustering of circulating endothelial cells. A. Global UMAP embedding of CD45⁺ immune and CD45⁻/CD31⁺ endothelial cells. Unsupervised clustering identified 12 distinct clusters corresponding to canonical immune populations and endothelial cells: 1. T CD4, 2. NKT-like, 3. Basophils, 4. B memory, 5. Monocytes, 6. NK, 7. T CD8, 8. ILCs, 9. B cells, 10. TCRγδ, 11. Endothelial cells and 12. Dendritic cells. These clusters correspond to well-established immune populations defined by canonical marker expression. B. Uniform Manifold Approximation and Projection (UMAP) of all patients based on FlowSOM unsupervised clustering of endothelial cell populations with bootstrapping. Each point represents a single endothelial cell, coloured by cluster assignment. C. Heatmap showing the abundance of endothelial markers across the identified clusters, highlighting differential marker expression patterns. D. Distribution of endothelial clusters across clinical categories, illustrating their association with distinct clinical phenotypes.

Association between immune and endothelial cell subsets and 90-day mortality

To explore differences in cellular subsets associated with survival, the relative abundances of immune and endothelial subsets were compared between deceased patients and 90-day survivors. Fig. 3 shows a volcano plot where each point represents a cellular cluster or subset. The X-axis depicts the median log-transformed difference in abundance between groups, and the Y-axis indicates statistical significance after multiple comparison correction (FDR, Benjamini-Hochberg method). Several immune and endothelial clusters were associated with mortality (FDR <0.05). The subsets showing associations linked to 90-day mortality were plasmablasts/memory IgG⁺, IgD⁻/IgM⁺ memory B cells and the CD36⁻/CD16⁺ endothelial cluster.

Fig. 3.

Volcano plot. Showing differential expression of immune and endothelial cell clusters between patients who died and those who survived at 90 days. Each point represents a cluster, with the x-axis indicating the median difference in log10-transformed cluster abundance and the y-axis showing the statistical significance as -log10 (FDR-adjusted p-value) adjusted by Benjamini-Hochberg correction. Clusters with FDR <0.05 are highlighted in red.

Prognostic value of immuno-endothelial subsets in 90-day mortality

To validate the ability of these three selected cell subpopulations to stratify patients, PCA was performed, as depicted in Fig. 4A. Each point represents a patient, coloured according to a cluster assignment. The plot shows a separation between groups, assessing the ability to stratify patients of these clusters as outcome predictors. Supplementary Table S2 summarises the distribution of patients across clusters and their 90-day outcomes. Clusters 2 and 3 predominantly comprised septic patients, with Cluster 3 enriched in septic shock cases. Notably, Cluster 2 exhibited the highest 90-day mortality (19.3%) compared to Clusters 1 (6.8%) and 3 (15.4%) (Fisher test p = 0.03455; two-sided p = 0.0498). To identify immuno-endothelial subsets with prognostic significance, univariate Cox analysis was performed for all clusters. Those associated with 90-day mortality were further refined using penalised LASSO-Cox regression to minimise overfitting. A recursive feature selection algorithm was used to identify the optimal multivariate model, which was restricted to subpopulations common to both analyses. The final model included three clinically relevant populations, adjusted for age and sex. Fig. 4B shows the three selected subsets, plasmablasts/memory IgG⁺, IgD⁻/IgM⁺ memory B cells and the CD36⁻/CD16⁺ endothelial cluster, all of which showed associations 90-day mortality. Immune clusters showed hazard ratios (HRs) > 2, while the CD36⁻/CD16⁺ cluster had a modest but significant effect (HR 1.64; p = 0.046). The model showed statistical significance (log-rank p < 0.001) and concordance (C-index = 0.73).

Fig. 4.

Clusters associated with mortality at 90 days. A. Principal component analysis (PCA) of patients using the expression of the three immune clusters selected by both univariate Cox regression and Lasso penalised Cox regression, followed by model refinement with stepwise AIC selection. Each point represents a patient, coloured by cluster assignment. B. Multivariate Cox proportional hazards model including the three selected immune clusters, adjusted for age and sex. The forest plot shows the hazard ratios (HRs) with 95% confidence intervals. All variables were scaled prior to modelling.

Multivariate risk modelling based on immuno-endothelial clusters associated with mortality

To explore the clinical utility of the immune and endothelial subsets identified as being associated with 90-day mortality, a risk score model was constructed based on the combination of the three previously selected cellular clusters. The risk score allowed patients to be stratified into three distinct risk groups: low, intermediate and high. As shown in Fig. 5A, patients were ranked according to their risk score, and a higher mortality was observed in the high-risk group (red) than in the intermediate (dark purple) and low (violet) groups. Fig. 5B shows this distribution, where the number of deaths progressively increased with the risk level, being especially high in the high-risk group (20/73) compared to the intermediate and low groups (6/73 each), reaching statistical significance (Fisher's exact test, p = 0.002). Finally, survival analysis using Kaplan-Meier curves (Fig. 5C) showed significant differences among the three risk groups (log-rank test, p = 0.0009). Supplementary Table S3 shows the distribution of patients across high- and low-risk groups according to the three selected immune/endothelial populations. Cluster 1 includes only low-risk patients, Cluster 2 concentrates most high-risk cases and Cluster 3 shows an intermediate profile. The association is significant (Fisher's exact test p = 1.41 × 10⁻¹¹), supporting internal consistency of the risk stratification.

Fig. 5.

Risk score stratifies 90-day mortality in three groups. A. Risk score distribution across patients, ranked from lowest to highest. Each patient is coloured by risk group (low, intermediate, high) and 90-day mortality status (alive or dead). Stacked bar plot showing the number of deaths per risk group: 6/67 in the low-risk group, 6/67 in the intermediate-risk group, and 20/53 in the high-risk group. The difference in mortality distribution was statistically significant (Fisher's exact test p = 0.002), B. C, Kaplan-Meier survival curves for the three risk groups, with number at risk shown below the plot. Survival significantly differed across the three groups (log-rank test, p < 0.01).

Evaluation of the immuno-endothelial model with SOFA and APACHE II in predicting mortality

To evaluate the discriminatory capacity of the proposed risk scoring model, its performance was compared with that of two widely used clinical scales in critically ill patients, SOFA and APACHE II. Supplementary Figure S6 shows the ROC curves corresponding to each model in the validation cohort. The cell population-based risk score (purple line) obtained an AUC of 0.931 (95% CI: 0.865–1.000), compared with both the SOFA (green) (AUC = 0.751; 95% CI: 0.600–0.903) and APACHE II scores (orange) (AUC = 0.804; 95% CI: 0.680–0.929), detailed in Supplementary Table S4. Supplementary Table S5 presents the confusion matrices and derived performance metrics for each model. The cell population-based risk score showed higher performance, with sensitivity (1.000), specificity (0.918) and balanced accuracy (0.959) compared with SOFA and APACHE II. Finally, Supplementary Table S6 reports the statistical comparison of ROC curves using DeLong's test and the cell population-based risk score discriminates mortality significantly better than SOFA (p = 0.0412) and APACHE II (p = 0.0408).

Identification and characterisation of CD36⁻/CD16+ endothelial cluster

The gating process applied to identify the CD36⁻/CD16⁺ cluster from viable events, singlets and ECs (CD45⁻/CD31⁺) is shown in Supplementary Figure S7A. To validate the correspondence between endothelial populations obtained by supervised and unsupervised analyses, classification analysis was performed by comparing the frequency of the CD36⁻/CD16⁺ clusters identified by both strategies. Supplementary Figure S7B indicates a statistically significant classification (Spearman's ρ = 0.82, p = 1.4 × 10⁻⁵⁴) between the cluster proportions identified by manual gating and those derived from the unsupervised analysis, confirming reproducibility of the unsupervised approach. Furthermore, Supplementary Figure S7C shows the differences in the proportion of events between the non-sepsis and sepsis groups, a progressive increase in patients with sepsis compared to the non-sepsis group. To further support endothelial lineage specificity and exclude potential immune contamination, we applied an alternative gating strategy incorporating CD146 as a complementary endothelial marker, yielding a highly concordant CD45⁻ endothelial population. In addition, no CD45 expression was detected within the endothelial cluster, supporting the absence of leucocyte contamination (Supplementary Figure S8).

Transcriptomic validation of endothelial and immune clusters by scRNA-seq

To validate our findings, we analysed publicly available single-cell RNA-seq data from patients with sepsis. Prior to clustering, ECs were lineage-restricted by selecting cells expressing CD31 and lacking CD45 expression, ensuring that downstream analyses were confined to bona fide endothelial populations. ECs were identified and clustered using Seurat, supporting the transcriptional heterogeneity observed in our flow cytometry data. This confirmed the presence of a dysfunctional endothelial subset, cluster 2 (CD36⁻/CD16⁺), which is associated with disease severity and poor outcomes. Fig. 6A shows the cluster distribution by clinical category, highlighting CD16⁺ (red), CD36⁺ (green) and cluster 2 (brown) cells. Fig. 6B depicts the progressive enrichment of cluster 2 in severe cases (sepsis and urological sepsis).

Fig. 6.

Transcriptomic validation of endothelial and immune clusters by scRNA-seq. A. Distribution of endothelial clusters across clinical categories in a public scRNA-seq. B. Proportion of each endothelial cluster in healthy, non-sepsis, sepsis and urinary tract infections sepsis group. C, Proportion of the CD16⁺/CD36^- endothelial cluster across clinical groups. D. Relative abundance of plasmablasts/memory IgG⁺ B cells across clinical groups. E. Relative abundance of IgD⁺/IgM⁺ memory B cells across clinical groups. F. Relative abundance of Monocyte MS1 cells across clinical groups. Statistical comparisons were performed using the Kruskal–Wallis test with post hoc multiple-comparison correction. Boxplots represent median and interquartile range; dots indicate individual samples.

Fig. 6C illustrates the distribution of endothelial cluster CD36⁻/CD16⁺, across clinical categories (healthy controls, non-septic critically ill patients, sepsis and urological sepsis). Consistent with our cytometry findings, the relative abundance of this endothelial population progressively increased with disease severity, being particularly enriched in patients with urological sepsis (Kruskal–Wallis, p = 8 × 10⁻⁴). In parallel, immune populations previously associated with sepsis-related immune dysregulation also displayed marked alterations. Plasmablasts/memory IgG⁺ B cells were significantly reduced in septic groups (Fig. 6D; Kruskal–Wallis, p = 3.3 × 10⁻⁴), as were IgD⁺/IgM⁺ memory B cells (Fig. 6E; p = 0.011). Conversely, the Monocyte MS1 subset, previously associated with sepsis diagnosis in the DUOS cohort, showed progressive enrichment across disease severity categories, reaching the highest levels in urological sepsis (Fig. 6F; p = 9.9 × 10⁻⁶). To determine whether the integrated endothelial-immune score, calculated using the same methodology as in the flow cytometry cohort, improves upon the previously reported Monocyte MS1 subset in the DUOS validation cohort, we compared their association with disease severity and mortality-related outcomes. The integrated risk score showed a positive correlation with disease severity as measured by the SOFA score (R = 0.63, p = 5.3 × 10⁻⁸), showing an association comparable to or greater than that observed for MS1 monocytes (R = 0.40, p = 0.0014) (Fig. 7A). Furthermore, the risk score was correlated with MS1 abundance (R = 0.60, p = 2.5 × 10⁻⁷). We next assessed the ability of both approaches to discriminate clinically relevant outcomes. Patients who died during hospitalisation showed significantly MS1 levels (p = 0.024) and risk scores (p = 0.03) compared with survivors (Fig. 7B). However, for 1-year mortality, only the integrated risk score remained significantly associated with outcome (p = 2 × 10⁻⁴), whereas MS1 alone failed to discriminate long-term mortality (p = 0.23). Patient stratification according to the integrated risk score further demonstrated progressive enrichment of in-hospital mortality among high-risk individuals (Fig. 7C). Consistently, ROC analysis showed predictive performance of the integrated score for mortality compared with MS1 alone, achieving an AUC of 0.877 (95% CI: 0.782–0.972) vs. 0.628 (95% CI: 0.419–0.838), respectively (Fig. 7D). Comparison of correlated ROC curves using DeLong's test showed a statistically significant difference between AUCs (DeLong's test: Z = 2.59, p = 0.0095).

Fig. 7.

Comparative performance of the integrated endothelial-immune risk score and the Monocyte MS1 subset in the DUOS validation cohort. A, correlation analyses between SOFA score, Monocyte MS1 abundance and the integrated risk score. B, comparison of MS1 monocyte abundance and integrated risk score according to in-hospital mortality and 1-year mortality. Statistical comparisons were performed using the Wilcoxon rank-sum test. C, distribution of patients ranked according to the integrated risk score, highlighting in-hospital mortality events. D, Receiver operating characteristic (ROC) curves comparing the ability of the integrated risk score and Monocyte MS1 abundance to discriminate mortality outcomes (AUC 0.877, 95% CI: 0.782–0.972) compared with MS1 alone (AUC 0.628, 95% CI: 0.419–0.838).

To further characterise the biological programs associated with the dysfunctional CD16⁺/CD36^- endothelial cluster, differential gene expression and pathway enrichment analyses were performed. Differential expression analysis revealed a transcriptional program enriched in inflammatory, endothelial activation and prothrombotic pathways (Supplementary Figure S9A). Functional enrichment analysis demonstrated overrepresentation of pathways related to leucocyte adhesion, cytokine signalling, and endothelial activation in Gene Ontology (GO) analyses (Supplementary Figure S9B), as well as vascular dysfunction and immune-endothelial signalling in KEGG pathway analysis (Supplementary Figure S9C). Reactome analysis further highlighted enrichment of inflammatory and coagulation-related pathways, consistent with a dysfunctional endothelial phenotype observed in severe sepsis (Supplementary Figure S8D).

Discussion

In this study, we applied a multidimensional approach based on high dimensional flow cytometry and unsupervised analysis to characterise the immune-endothelial profile of critically ill postoperative patients with sepsis and SS. Our analysis yielded three key findings: First, we identified immune and ECs clusters and subsets associated with 90-day mortality, both in univariate analyses and through multivariate selection using LASSO-Cox regression. Second, we developed a risk score model integrating three of these subsets, two immune (plasmablasts/IgG⁺ memory and IgD⁻/IgM⁺ memory) and one endothelial cluster (CD36⁻/CD16⁺), which allowed patient stratification into three groups with different survival outcomes. The model showed promising predictive performance and outperformed conventional clinical scores such as SOFA and APACHE II. Lastly, among the identified clusters, the CD36⁻/CD16⁺ cluster was consistently associated with adverse outcomes, suggesting a dysfunctional endothelial phenotype relevant to the clinical progression of sepsis.

Sepsis is characterised by profound immune dysfunction and endothelial damage. At the immunological level, plasmablast expansion and alterations in memory B cells have been previously described and are associated with immunosuppression and severity prognosis.²⁷ Endothelial activation and dysfunction have also been documented, but detailed phenotypic characterisation of circulating endothelial subpopulations remains limited. Although previous studies using multiparametric flow cytometry identified immune subpopulations associated with mortality in sepsis,^28,29 the simultaneous integration of immune and endothelial profiles into a prognostic model remains less explored.

Moreover, most studies have evaluated immune and endothelial biomarkers separately and have predominantly relied on soluble circulating mediators rather than cellular phenotyping approaches.^30,31 Traditional biomarkers of endothelial activation and injury, including angiopoietin-2 (Ang-2), soluble thrombomodulin (sTM), VWF, soluble intercellular adhesion molecule 1 (sICAM-1), soluble vascular cell adhesion molecule 1 (sVCAM-1), E-selectin, endocan and syndecan-1, have consistently been associated with vascular dysfunction, organ failure and mortality in sepsis.^32,33 Likewise, immune-related biomarkers such as IL-6, soluble TNF receptors, monocyte HLA-DR suppression, and lymphocyte dysregulation have shown prognostic value but incompletely capture the complexity of host response heterogeneity.³⁴

In our prospective cohort of 219 critically ill surgical patients, we identified two B cell subpopulations (plasmablasts/IgG⁺ and IgD⁻/IgM⁺ memory B cells) and one specific endothelial cluster, CD36⁻/CD16⁺, as associated with 90-day mortality. Plasmablast expansion may indicate ongoing immune activation, whereas alterations in IgD⁻/IgM⁺ memory subsets could reflect ineffective immune responses.^1,35 The absence of CD36 and coexpression of CD16 in ECs suggests an activated, proinflammatory phenotype,^36,37 consistent with loss of homoeostatic function and amplification of inflammation. These endothelial changes may contribute to multiorgan dysfunction, representing a subset of inflammatory ECs capable of enhancing innate immune responses.³⁸ In contrast to conventional soluble biomarkers, our high-dimensional cellular approach may capture functional immune-endothelial heterogeneity at the single-cell level, potentially providing a more mechanistic understanding of host dysregulation in sepsis. Therefore, our findings extend current knowledge by integrating both immune and endothelial cellular states into a single prognostic framework rather than evaluating each compartment independently.

Although SOFA and APACHE II scores are general assessment tools, they lack sensitivity to detect specific cellular processes in the host.^39,40 Our immuno-endothelial risk score, integrating two immune and one endothelial population, showed improved predictive performance (AUC 0.935 vs. 0.751–0.804) compared to these clinical indices. This underscores the potential added value of cellular biomarkers for early and precise risk stratification, in line with recent studies advocating their inclusion in prognostic models.⁴¹

Importantly, these findings should be interpreted in the context of current sepsis biomarker signals used for prognostic benchmarking. Conventional biomarkers such as C-reactive protein (CRP) and procalcitonin (PCT) remain clinically useful for diagnosis and monitoring but show limited prognostic specificity.³⁰ Similarly, endothelial and immune biomarkers including Ang-2, sTM, syndecan-1, IL-6 and HLA-DR have shown associations with mortality and organ dysfunction but individually provide only partial information on the biological heterogeneity of sepsis.⁴² In this context, our integrated cellular signature may offer complementary prognostic value by simultaneously capturing immune dysregulation and endothelial dysfunction within the same model.

To validate our results, we analysed publicly available single-cell RNA sequencing (scRNA-seq) data from patients with sepsis. Endothelial clustering confirmed transcriptional heterogeneity consistent with our cytometric findings.⁴³ A specific endothelial subset was enriched in severe cases and matched the CD36⁻/CD16⁺ phenotype. Functional enrichment analyses supported this correspondence: GO terms involved leucocyte adhesion and cytokine signalling,⁴⁴ KEGG pathways highlighted vascular dysfunction, and Reactome revealed the activation of proinflammatory and coagulation cascades. The integration of immune and endothelial signatures demonstrated overlap between scRNA-seq and cytometry, reinforcing the relevance of this dysfunctional endothelial subset as a prognostic biomarker.³⁸ Future studies should aim to characterise the origin and functional role of these populations to better understand their contribution to disease progression. In addition to endothelial alterations, our validation analyses confirmed the relevance of the Monocyte MS1 subset, a transcriptionally defined monocyte program previously identified as a hallmark of bacterial sepsis through single-cell transcriptomic profiling. Originally, immature CD14⁺ monocyte state characterised by immunosuppressive features, reduced antigen-presentation capacity and expansion in severe infection, MS1 monocytes have emerged as a key component of sepsis-associated immune dysregulation.⁴³ Subsequent studies further demonstrated that increased MS1 abundance correlates with greater disease severity, shock and broad host-response perturbation, supporting their relevance as biomarkers of sepsis pathophysiology rather than merely diagnostic signatures.⁴⁵ More broadly, monocytes are increasingly recognised as central orchestrators of sepsis, undergoing profound functional reprogramming from early inflammatory activation toward dysfunctional or immunosuppressive states that may contribute to persistent immune dysregulation and adverse outcomes.⁴⁶ In our validation cohort, MS1 monocytes were associated with disease severity and in-hospital mortality, consistent with these previous observations. However, the integrated endothelial-immune score showed associations with organ dysfunction and long-term outcomes and outperformed MS1 alone in mortality discrimination. These findings suggest that while MS1 captures an important dimension of sepsis-induced immune dysfunction, integrating endothelial and immune cellular states may provide a more comprehensive representation of host-response heterogeneity and improve prognostic stratification.

Our findings have several important implications. The integration of immune and endothelial biomarkers provides improved prognostic accuracy compared with traditional scales, enabling the early identification of high-risk patients and potentially guiding personalised interventions. The identified subsets of activated ECs, plasmablasts and altered memory B cells represent promising therapeutic targets for restoring immune-endothelial homoeostasis. This study identifies a previously undescribed endothelial cluster with prognostic value and supports the concept that integrated immune-endothelial profiling may represent a clinically relevant approach to precision medicine in postoperative sepsis. This cellular score could complement the current severity scales and support personalised monitoring, resource allocation and patient selection for clinical trials.

The main limitation of this study is the lack of longitudinal data on biomarker dynamics, which prevents the assessment of their temporal evolution. One notable strength is the use of a homogeneous cohort, which minimises variability and allows for a more precise interpretation of the results. Nevertheless, because the cohort included mostly post-surgical patients with abdominal sepsis, the applicability of our findings to other sepsis types remains to be determined. Another limitation is that neutrophils were not represented in either the derivation flow cytometry dataset or the validation scRNA-seq dataset, as both analyses primarily focused on PBMC populations. Considering the well-established role of neutrophils in sepsis pathophysiology, their absence limits the comprehensiveness of the immune-endothelial landscape described herein. Further validation in independent cohorts and functional studies is required to elucidate the mechanistic role of these subsets. Functional studies, both in vitro and in vivo, are needed to confirm the active role of the identified cell subpopulations.

In conclusion, this study suggests that an integrated immune-endothelial cellular signature may improve risk stratification and is associated with 90-day mortality in critically ill patients with sepsis, offering a promising tool for precision prognosis and guiding future therapeutic trials.

Contributors

Conceptualisation: RD-PU, AG-C and ET. Methodology: RD-PU, AG-C, ET, DB and MP-M. Formal Analysis: RD-PU, AG-C. Investigation: DB, AG-C, ET, MP-M, MB-C, EG-S, RC-Z, IR-M, JM-I, LS-DP, MM-F, MT-D, PA, EC-S, SC-C, AF-U, MF-A, JM-E, IS-M, HG-B, RL-H, GA, RH, CI and RP-A. Validation: RD-PU and AG-C. Writing: Original Draft Preparation: RD-PU and AG-C. Writing: Review & Editing: AT-V and ET. Visualisation: RD-PU, AG-C and ET. Supervision: AG-C, ET and DB. Project Administration: AG-C and ET Funding Acquisition: ET and AG-C. All authors read and approved the final manuscript.

Data sharing statement

The dataset generated and analysed during the current study is not publicly available due to patient privacy and data protection regulations. Access to de-identified individual participant data may be made available from the corresponding author upon reasonable request, subject to approval by the institutional ethics committee and in accordance with applicable data protection legislation. Requests for data access should include a methodologically sound proposal and will be evaluated on a case-by-case basis. Data will be shared under a data access agreement for non-commercial research purposes only.

Declaration of interests

All authors declare that they have no conflicts interest. None of the clinical investigators received an honorarium for participating in the study.