An interpretable clinical-proteomic framework identifies baseline neuroinflammatory proteomic patterns associated with post-thrombectomy mortality

Abstract

Endovascular thrombectomy (EVT) is the standard reperfusion therapy for emergent large-vessel occlusion; however, nearly 20% of EVT-treated patients still die within one year. Although early clinical indicators can help stratify post-EVT prognosis, the baseline neuroinflammatory biology associated with fatal outcomes after EVT remains incompletely characterized. In this prospective study of 148 EVT-treated patients, including 30 patients who died within 1 year, we profiled pre-procedure serum proteins using the Olink 384 Inflammation panel and applied an exploratory clinical-proteomic framework. A clinical-only model was used to define higher-risk and lower-risk clinical strata. We then mapped these strata to baseline proteomic patterns through differential expression and annotation-supported candidate mapping. Higher-risk phenotypes were associated with a candidate inflammation- and immune-receptor-related baseline protein pattern. Lower-risk phenotypes were associated with a candidate adhesion/extracellular-matrix-related baseline protein pattern. These internally derived findings suggest that poor outcomes after EVT are associated with candidate baseline inflammatory and tissue-structure-related proteomic patterns. This study provides a neuroinflammation-focused framework for hypothesis generation and candidate prioritization, but the findings require external validation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12974-026-03971-4.

Abbreviated summary

Jixian Lin et al. examine pre-procedure serum proteins in patients with large-vessel stroke and report candidate baseline protein patterns associated with subsequent mortality risk. Higher-risk patients showed candidate baseline protein patterns related to inflammation and immune signaling, whereas lower-risk patients showed patterns related to tissue adhesion and extracellular-matrix organization.

Introduction

Acute ischemic stroke due to emergent large-vessel occlusion (ELVO) remains a leading cause of long-term disability and mortality worldwide [1]. Endovascular thrombectomy (EVT) has been firmly established as the standard of care for restoring cerebral perfusion in ELVO [2–4]. Yet, successful recanalization alone does not ensure favorable neurological recovery: a substantial subset of patients, even after successful reperfusion, experience limited functional recovery or markedly increased 1-year mortality [5–7]. This prognostic heterogeneity underscores an urgent need to move beyond vessel patency and identify patients who remain at high risk for poor outcomes after EVT, despite successful recanalization [8, 9].

In this context, prognostic models based on routinely available early clinical indicators remain the most practical and widely used tools, including age, National Institutes of Health Stroke Scale (NIHSS) score, comorbidities, imaging findings, treatment timing, and early laboratory measures [10–13]. These models are inexpensive, readily accessible in hospital settings, and have demonstrated predictive utility [14–16]. However, although clinical scores can identify high-risk patients, they do not explain the biological processes linking clinical variables to adverse outcomes. Even when general interpretability algorithms such as SHapley Additive exPlanations are applied, their explanations typically remain limited to feature importance rather than providing biologically interpretable, pathway-level insight [17–20].

To address this biological gap, recent studies have increasingly turned to high-throughput omics approaches, particularly serum proteomics and inflammation-related biomarker profiling [21–25]. Proteomics offers a window into the dynamic functional state of disease, capturing processes such as neuroinflammation [21, 22], blood-brain barrier injury and repair, microglial activation, and endogenous neurorestorative responses after stroke [23–25]. However, translating molecular complexity into clinically interpretable representations remains challenging. To bridge molecular information with clinically meaningful phenotypes, we used a knowledge-distillation-inspired exploratory framework in which an out-of-fold (OOF) clinical-only risk phenotype (ClinIndex) served as an intermediate representation linking early post-EVT risk profiles with baseline proteomic patterns. This design preserved the biological depth of proteomics while providing a clinically transparent representation of disease-associated biology [26–28]. Through this framework, our study aimed to characterize baseline proteomic patterns associated with post-thrombectomy mortality and to prioritize candidate signatures for future studies of maladaptive inflammation, endothelial dysfunction, and impaired neurorepair.

Methods

Study design and patient population

This prospective, single-center observational study (ChiCTR2100045753) consecutively enrolled 148 patients with acute ischemic stroke due to emergent large-vessel occlusion (ELVO) who underwent endovascular thrombectomy (EVT) at Minhang Hospital, Fudan University, between January 2023 and June 2024. The protocol was approved by the local Institutional Review Board, and written informed consent was obtained from all participants or their legally authorized representatives.

Eligibility criteria included age 40–100 years, presentation within 24 h of onset, and moderate-to-severe large-vessel occlusion stroke, defined as an admission National Institutes of Health Stroke Scale (NIHSS) score of 4–22. Key exclusion criteria were active hemorrhage, severe coexisting systemic illness, or malignancy with < 3 months life expectancy. Stroke etiology was classified according to the Trial of Org 10,172 in Acute Stroke Treatment (TOAST) criteria, including large-artery atherosclerosis (LAA) and cardioembolism (CES); one patient with cancer-associated embolism was excluded from etiology-specific comparisons. Additionally, 26 age- and sex-matched healthy controls (HCs) were recruited solely as a baseline reference for proteomic analyses and were excluded from clinical outcome comparisons. Controls were recruited from hospital health examination participants or community volunteers during the same study period and had no history of acute stroke or active inflammatory, malignant, or severe systemic disease.

Clinical, laboratory, and outcome assessment

We prospectively collected baseline demographics (age and sex), vascular risk factors, major comorbidities (hypertension, diabetes, dyslipidemia, coronary/cardiac disease, renal dysfunction, and prior stroke), pre-stroke functional status assessed by the modified Rankin Scale (pre-stroke mRS), occlusion location, baseline infarct burden assessed by the Alberta Stroke Program Early CT Score (ASPECTS), treatment timing variables (including onset/last-known-normal-to-arrival, puncture-to-first-angiography, and puncture-to-recanalization intervals), and concomitant acute treatments (including intravenous thrombolysis and reperfusion-therapy modality). Neurological severity was serially evaluated using NIHSS at admission, post-procedure, 24 h, and 7 days. Key procedural data (e.g., EVT passes and reperfusion grade) were systematically documented. Reperfusion was assessed using the modified Thrombolysis in Cerebral Infarction (mTICI) scale in the source dataset; successful reperfusion was defined as mTICI ≥2b. Because this cohort was collected during a transition period before routine expanded Thrombolysis in Cerebral Infarction (eTICI) implementation, mTICI was retained for primary analyses to preserve consistency with prospectively recorded variables. Functional status was evaluated using mRS at discharge and 1-year. The primary endpoint was all-cause mortality at 1 year. Additional available clinical, EVT, timing, imaging, and laboratory indicators are summarized in Supplementary Tables S1 and S2. The secondary functional endpoint was unfavorable outcome at 1 year (mRS 4–6).

A comprehensive laboratory panel was measured pre-EVT and 24–48 h post-procedure. This included metabolic markers (glucose, hemoglobin A1c [HbA1c], total cholesterol, triglycerides, low-density lipoprotein cholesterol [LDL-C], high-density lipoprotein cholesterol [HDL-C], and homocysteine), inflammatory markers (high-sensitivity C-reactive protein [hs-CRP], leukocyte count, neutrophil count, and procalcitonin), coagulation markers (fibrinogen, platelets, D-dimer, prothrombin time/international normalized ratio [PT/INR]), cardiac injury/stress markers (high-sensitivity troponin I [hs-cTnI] and B-type natriuretic peptide [BNP]), and renal function markers (creatinine, blood urea nitrogen [BUN], and estimated glomerular filtration rate [eGFR]).

Sample collection and proteomic profiling

Peripheral blood was collected from all patients immediately before EVT and before endovascular device deployment. Samples were processed within 1 h: serum was separated (1,500 × g, 15 min, 4 °C), aliquoted, flash-frozen, and stored at − 80 °C. Samples were randomized and shipped on dry ice for analysis.

Serum proteins were quantified using the Olink Target 384 Inflammation panel (Olink Proteomics, proximity extension assay technology), following the manufacturer’s protocol. Protein expression was reported as log₂-scaled Normalized Protein eXpression (NPX). For quality control using OlinkAnalyze (v4.3.1), values below the limit of detection were treated as missing. Proteins with > 50% missing data were excluded, retaining 345 proteins for analysis. For predictive modeling, residual missing values were imputed using protein-wise medians estimated within each training fold and then applied to the corresponding held-out folds to avoid information leakage. For descriptive visualizations, median imputation was performed on the baseline NPX matrix without using outcome labels.

Statistical and interpretable analysis framework

Our analytical framework followed the multilayer workflow illustrated in Fig. 1A, which summarizes the cohort, data layers, and integrated analysis roadmap. The panel summarizes the prospective ELVO cohort, pre-EVT serum collection, EVT procedure, 1-year follow-up, and outcome distribution, separates the clinical/laboratory data layer from the pre-EVT proteomic layer, and outlines the integrated roadmap: clinical-risk-factor screening, mortality proteomic signal, clinical-proteomic bridge, domain-domain interaction (DDI) functional organization, and protein-domain candidate prioritization. Briefly, clinical-risk-factor screening used early post-EVT clinical and laboratory variables to generate strict out-of-fold (OOF) mortality probabilities and exploratory ClinIndex High and ClinIndex Low strata. The mortality proteomic signal layer characterized pre-EVT Olink proteins associated with 1-year mortality. Clinical-proteomic bridging related the clinical risk phenotype to baseline proteomic and co-expression structure through ClinDistill and weighted gene co-expression network analysis (WGCNA). The domain-domain interaction layer used protein-protein interaction (PPI)/domain-domain interaction (DDI) annotations to organize candidate network and domain-level patterns. Protein-domain candidate prioritization then used bridge-protein, path-scoring, four-ring, and protein-domain-domain-protein (P-D-D-P) visualizations for annotation-supported prioritization only, not for causal inference, mediation testing, or clinical validation. This sequence reduces direct circular construction but does not eliminate same-cohort internal dependence: ClinIndex is generated upstream from clinical-only OOF predictions, whereas ClinDistill, WGCNA, PPI/DDI, and path-scoring layers are downstream descriptive or prioritization analyses. Detailed abbreviations and source-variable terms are provided in Supplementary Table S3; PPI-DDI annotation records are provided in Supplementary Table S6. Methodological concepts and analytical rationale are summarized in Supplementary Table S9, sections S9.1 and S9.2, respectively.

Fig. 1.

Study design and exploratory clinical-proteomic analysis overview. A Study design and analysis roadmap. The panel summarizes the prospective ELVO cohort (n = 148), healthy controls used as a baseline reference (HCs; n = 26), pre-EVT serum collection, EVT procedure, 1-year follow-up, mortality distribution, clinical/laboratory data layers, pre-EVT Olink proteomics, and downstream clinical-proteomic analyses. B Point-biserial correlation coefficients (r_pb) for the top 30 clinical indicators associated with 1-year mortality. This panel represents descriptive screening only, not the final model-eligible predictor set. Red bars indicate positive associations with death; blue bars indicate negative associations. C Heatmap of the top 30 clinical indicators. Columns are annotated by outcome and clustered within outcome groups; rows are split by positive versus negative r_pb. Data are z-score normalized per feature and truncated at ±2.5. D Univariate spline regression using generalized linear models (df = 3) for six representative indicators (7-Day NIHSS, Age, Diff-PCT, Post-D-dimer, Post-hs-cTnI, and Post-NSE) versus predicted probability of death. The solid line indicates the fitted curve, and the shaded area indicates the 95% confidence interval

Each analytical layer had a prespecified role. The clinical-only model defined an internally derived early post-EVT clinical-risk anchor. Differential proteomic analysis described proteins associated with 1-year survival status. ClinDistill examined whether baseline proteins could represent the clinical OOF risk signal in protein space. WGCNA summarized coordinated protein co-expression structure, whereas PPI/DDI annotation provided external database-supported biological context. Bridge-protein ranking and four-ring path scoring were used only to prioritize candidate patterns for visualization and future validation.

Clinical risk identification and differential proteomics

First, we identified clinical indicators associated with 1-year mortality using point-biserial correlation screening for the binary mortality endpoint, with Benjamini-Hochberg false discovery rate (BH-FDR) correction. Generalized linear models (GLMs) with natural splines (degrees of freedom [df] = 3) assessed nonlinear risk trends.

Second, we identified differentially expressed proteins (DEPs) using the R limma package (v3.62.2), with the main exploratory DEP set defined using nominal P < 0.05. The primary limma design matrix included outcome group and prespecified covariates to adjust for potential confounding by age, pre-EVT NIHSS, and mTICI reperfusion grade. An unadjusted sensitivity analysis was also performed to assess the robustness of key DEP signals. Benjamini-Hochberg false discovery rate (BH-FDR)-adjusted q values were calculated for all tested proteins, and BH-FDR-supported proteins were defined as q < 0.05 and reported in Supplementary Table S4. The primary exploratory nominal DEPs were then used for principal component analysis (PCA) and heatmap clustering. To complement this, variable importance in projection (VIP) scores were computed from a partial least-squares model (pls package v2.8-0) on scaled NPX features.

Model building, cross-layer association, and core networks

We used two model components with distinct roles: (i) a clinical-only model that generated strict OOF clinical risk probabilities and defined the exploratory ClinIndex strata; and (ii) a proteomic ClinDistill representation in which protein inputs learned the clinical OOF risk signal as a soft teacher target. Figure 3A provides an overview of the model workflow, whereas Fig. S3 provides the detailed distillation schematic: the clinical model was trained within each training fold, generated pclin only for the held-out fold, and this OOF-predicted clinical risk was then used as the soft teacher target for the protein-NPX student model; the student model minimized the distance between pdistill and pclin and yielded sparse protein coefficients for exploratory interpretation. Elastic net is a penalized regression approach that performs variable selection while reducing overfitting when predictors are numerous and correlated. The primary input screen excluded seven prespecified leakage variables (DC-mRS, 1-yr mRS, Pre-EVT NIHSS, Post-EVT NIHSS, 7-Day NIHSS, Pre-mRS, and 24 h-NIHSS) before model fitting. After these exclusions, 246 variables remained model-eligible. Outcome_Death_1 year was the target label and was not used as a candidate predictor. Baseline CT-ASPECTS was reported descriptively and evaluated in a post hoc sensitivity audit using the same preprocessing, folds, screening, and tuning procedure; complete quantitative pre-treatment infarct-core volume was unavailable. Predictors were restricted to variables available within the early post-EVT window (through day 7); no post-discharge or outcome-derived variables were entered. Clinical and laboratory predictors were imputed and standardized within each training fold, with fold-specific preprocessing parameters applied to the held-out fold to avoid information leakage. Internal performance was summarized using 5-fold OOF predictions, with 5-fold cross-validation selected to balance bias and variance. ClinDistill did not define ClinIndex strata and was not used as independent validation of ClinIndex; it was retained as a downstream protein-space representation for exploratory interpretation.

Fig. 3.

Clinical-risk modeling, ClinDistill workflow, cross-layer association, and core subgroup PPI context. A Modeling workflow panel. The panel separates pre-/post-EVT variable availability, the 246 model-eligible clinical inputs after exclusion of seven leakage variables, clinical-only elastic-net (ENet) out-of-fold (OOF) teacher modeling, ClinIndex High/Low stratification, pre-EVT serum proteomics, mortality-associated DEPs, WGCNA mapping, and the ClinDistill student representation with 18 core proteins (Core-DEP = ClinDistill-selected and differentially expressed; Core-non-DEP = ClinDistill-selected and not differentially expressed). B OOF ROC comparison of the clinical-only model (AUC = 0.825) and ClinDistill (AUC = 0.824). AUCs represent internal OOF performance and should not be interpreted as external validation or model superiority. C Non-zero coefficient summaries for the clinical-only model and ClinDistill. Red bars indicate positive mortality-associated coefficients, whereas blue bars indicate negative mortality-associated coefficients. These coefficients are descriptive and do not establish causal biological effects. D Cross-layer Spearman correlation heatmaps among selected clinical predictors, ClinDistill-derived core proteins, and WGCNA module eigengenes. Asterisks denote significance as shown in the figure; these panels summarize associations rather than mediation. * P < 0.05; ** P < 0.01; *** P < 0.001. E PPI network context for Core-DEP and Core-non-DEP proteins. Node size reflects degree, node color indicates subgroup assignment, and edge width reflects available DDI count where applicable. These networks are annotation-supported summaries

To bridge clinical and proteomic data, we used weighted gene co-expression network analysis (WGCNA) module eigengenes (MEs) to compute Spearman correlations (Clin × ME, Prot × ME, and Clin × Prot), presented as a Clin→ME→Prot association heatmap. Finally, we explored the internal PPI networks (STRING database) of the resulting core protein subgroups, which were categorized according to whether they overlapped with DEPs (Core-DEP) or did not overlap with DEPs (Core-non-DEP).

Exploratory functional annotation of core subgroups

To systematically characterize the core subgroups, we performed a multiscale analysis. We visualized the Core-DEP PPI network, profiled its functional class composition using a built-in domain dictionary, and plotted frequency counts. Separately, we performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment analyses using clusterProfiler v4.14.6. At a macro level, we compared functional preferences by mapping PPIs to functional class pairs (ClassA-ClassB) and visualizing aggregated interaction patterns for Core-DEP and Core-non-DEP groups using chord diagrams. At the micro level, we performed domain-domain interaction (DDI) enrichment using the protein-protein interaction/domain-domain interaction mapping (PPI-ID) platform, which annotates structural domain pairs. We then plotted the most frequent DDI pairs for the Core-DEP and Core-non-DEP groups. The complete PPI-DDI annotation records used for these annotations are provided in Supplementary Table S6. Briefly, the batch-processing script batch_process_ppi.R retrieved InterPro/Pfam domain annotations for UniProt accessions and matched candidate domain pairs against compiled PPI-ID domain-interaction resources. The resulting outputs were standardized and compiled into Supplementary Table S6, which contains protein-pair, UniProt, domain, and domain-range annotation records used by the downstream scripts that generated Figs. 4 and 5.

Fig. 4.

Protein subgroup definition and domain-level functional organization. A DDI annotation workflow panel. The five-step schematic summarizes the full measured protein panel (n=345), network filtering, retained network nodes, DDI network construction, and downstream annotation outputs. B DDI network across Core-DEP, Core-non-DEP, Other-DEP, and Other-non-DEP proteins. Node colors indicate protein groups, and edges summarize annotated DDI connectivity. C Group-specific domain-class count profiles for Core-DEP and Core-non-DEP linked records. Core-DEP-linked records show a broad repertoire led by Immunoglobulin, Adhesion/extracellular matrix (Adhesion/ECM), tumor necrosis factor receptor (TNFR), Protease/Coagulation, leucine-rich repeat (LRR), Other, and Cytokine/growth-factor (GF) ligand classes. Core-non-DEP-linked records are characterized by Adhesion/ECM, followed by Other, Immunoglobulin, Protease/Coagulation, LRR, and G protein-coupled receptor (GPCR) classes. D Compact domain-pair inventory for Core-DEP and Core-non-DEP proteins. The panel summarizes recurrent DDI pairings for Core-DEP and Core-non-DEP proteins and serves as the upstream domain-pair vocabulary for the path-level P-D-D-P candidate connections shown in Fig. 5D

Fig. 5.

Bridge-protein prioritization and candidate domain-domain annotation. A Candidate path-scoring schematic. The panel explains how ClinIndex Low/High strata are linked to protein-level standardized contrasts (Cohen's d), W1-W2-W3 segment weights, bottleneck path ranking, function-pair aggregation, and downstream domain-level interpretation. B Four-ring candidate path view linking ClinIndex strata, Core proteins, Other proteins, and outcome-annotation sectors. Outcome sectors are shown for contextual annotation only and are not used in path scoring. Ring sizes reflect sample counts (ClinIndex High = 54, ClinIndex Low = 94; Death = 30, Survivor = 118), and node/edge weights are within-cohort ranking metrics based on standardized effect-size summaries and DDI-supported connectivity. C Candidate connection quantification and aggregated functional-pair summary. Segment bars display ClinIndex-Core, Core-Other, and Other-Outcome adjacency weights; the functional-pair panel collapses Core-Other candidate connections into functional classes stratified by ClinIndex High and Low sectors. D Protein-domain-domain-protein (P-D-D-P) alluvial plots for selected candidate links. The lower-risk panel (ClinIndex Low) displays selected candidate connections with domain annotations. The higher-risk panel (ClinIndex High) displays selected candidate connections with domain annotations. Stream width reflects PathWeight_min

Integrated path visualization and candidate domain-domain annotation

Identifying key bridge proteins

In the final candidate-path synthesis, bridge proteins were defined as proteins with high intergroup connectivity linking core protein groups (Core-DEP or Core-non-DEP) to the Other-DEP layer in the external PPI graph. Connectivity was quantified as the unweighted number of cross-group interactions (cross-group degree). Proteins were ranked by this metric, and top-ranked proteins were prioritized as candidate connectors between groups.

Constructing the ClinIndex risk strata

Next, we constructed the ClinIndex by dichotomizing OOF prediction probabilities from the clinical-only model using the Youden index to stratify patients as ClinIndex High or ClinIndex Low. Using OOF predictions reduces in-sample optimism; therefore, these strata are exploratory. In this study, ClinIndex was used as an internally derived early post-EVT clinical-risk phenotype because selected clinical inputs included early dynamic variables. Figure 3A and Fig. S3 show the modeling sequence: clinical-only OOF predictions define ClinIndex strata, whereas ClinDistill learns this OOF risk signal downstream in protein space.

Assembling the four-ring path view and scoring principles

We then assembled a four-ring display linking ClinIndex strata, Core proteins, Other proteins, and outcome-annotation sectors, with ring sizes scaled to sample proportions. Protein node sizes were determined by endpoint strength: S_i = |d_i|, where d_i is the standardized effect size (Cohen’s d) of z-scored protein abundance between ClinIndex High and ClinIndex Low strata. Z-scoring was performed on NPX values before effect-size estimation.

Segment weights were defined as follows: (i) ClinIndex-Core weight W1 = SCore; (ii) Core-Other weight W2 = sqrt(|SCore| × |SOther|) (geometric mean of endpoint strengths); and (iii) Other-Outcome weight W3 = SOther. Complete ClinIndex-Core-Other-Outcome paths were scored using a bottleneck principle: PathWeight_min = min(W1, W2, W3).

Candidate path quantification and functional pattern summarization

We enumerated all ClinIndex-Core-Other-Outcome chains and ranked them by PathWeight_min. Top paths were visualized as stacked bars showing segment weights for ClinIndex-Core, Core-Other, and Other-Outcome adjacency components. To summarize functional patterns, we collapsed Core-Other candidate connections into domain pairs, mapped domains to functional classes, including tumor necrosis factor receptor/tumor necrosis factor (TNFR/TNF) and extracellular matrix/adhesion (ECM/Adhesion), and aggregated their contributions (weighted by PathWeight_min), stratified by outcome. The top N function pairs for each outcome were plotted.

In-depth candidate domain-domain annotation

Finally, for candidate domain-domain annotation, we visualized the domain-domain interaction (DDI) pairings associated with selected Core-Other protein-protein interaction (PPI) candidate connections. We decomposed these displayed connections into P-D-D-P (Core protein-Domain 1-Domain 2-Other protein) alluvial plots. Direct Core-Other domain annotations were used when available; when displayed domain pairs were traced from source annotation tables, the alluvial plot retained the Core protein, Domain 1, Domain 2, and Other protein layers for readability. Stream width reflected PathWeight_min and was stratified by outcome. The PPI-DDI annotation records provided the source records for the Fig. 4D domain-pair inventory, whereas the displayed P-D-D-P source records for Fig. 5D are provided separately in Supplementary Table S8.

Results

Post-EVT clinical risk phenotype is characterized by dynamic indicators and nonlinear mortality patterns

To establish the clinical and proteomic context for downstream analyses, we first characterized the clinical risk phenotype in our prospective ELVO cohort (n = 148; Supplementary Tables S1 and S2). Twenty-six age- and sex-matched healthy controls (HCs) were included solely as a baseline reference for proteomic analyses. As shown in Fig. 1A, the study followed patients from pre-EVT assessment and serum collection through EVT, post-EVT clinical evaluation, and 1-year outcome assessment. The integrated data framework included pre- and post-EVT clinical/laboratory variables and pre-EVT serum proteomic profiling using the Olink Target 384 Inflammation panel, with 345 proteins retained for analysis. At 1 year, all-cause mortality was 20.3% (30 deaths and 118 survivors). Reperfusion success, defined as modified Thrombolysis in Cerebral Infarction (mTICI) ≥2b, was achieved in 145/148 patients (98.0%), including 116/118 patients in the Survivor group (98.3%) and 29/30 patients in the Death group (96.7%). Intravenous thrombolysis was administered in 52/148 patients (35.1%), including 43/118 patients in the Survivor group (36.4%) and 9/30 patients in the Death group (30.0%). As a complementary functional endpoint, unfavorable outcome at 1 year (mRS 4–6) occurred in 61/148 patients (41.2%), including 31/118 patients in the Survivor group (26.3%) and all patients in the Death group. Thus, despite a very high reperfusion success rate, substantial residual mortality and functional disability persisted, motivating further analyses beyond vessel reopening alone.

We performed point-biserial correlation screening to identify clinical indicators most strongly associated with 1-year mortality (Fig. 1B). Established factors such as 7-Day NIHSS and Post-BUN were among the strongest mortality-associated indicators, while laboratory indicators, including Post-D-dimer and Diff-CRP, were also associated with mortality in this descriptive analysis. A heatmap of these top 30 indicators, with label definitions and source mappings provided in Supplementary Table S3, further supported outcome-related stratification, showing that variables positively associated with mortality were higher in the Death group and variables negatively associated with mortality were higher in the Survivor group (Fig. 1C). These screening analyses were descriptive and did not represent the final model-eligible predictor set.

To characterize nonlinear risk patterns, we further applied univariate spline regression (Fig. 1D). This analysis revealed marked nonlinear increases in predicted probability of death associated with Age, 7-Day NIHSS, Diff-PCT, Post-D-dimer, Post-hs-cTnI, and Post-NSE, with Age showing a notable risk increase after approximately 60 years. These nonlinear patterns suggested that post-EVT risk was not fully captured by simple linear trends alone, supporting the use of clinically informed exploratory stratification for subsequent clinical-proteomic analyses.

Baseline serum proteomic signatures are associated with 1-year survival status

We first examined the overall baseline inflammatory-protein landscape of EVT-treated patients. Compared with healthy controls, pre-EVT serum proteomic profiles showed broad inflammatory-protein alterations in the EVT cohort (Fig. S1A). Complementary enrichment and pairwise correlation analyses further highlighted immune activation, cytokine/chemokine signaling, leukocyte-related pathways, and coordinated inflammatory-protein correlation patterns (Fig. S2A-B).

After establishing the cohort-level inflammatory-protein background, we next examined whether the clinical risk phenotype was accompanied by baseline molecular differences according to 1-year survival status. Differential proteomic analysis comparing the Death and Survivor groups identified 61 nominal mortality-associated DEPs adjusted for age, pre-EVT NIHSS, and mTICI, including 52 proteins higher in the Death group and 9 proteins higher in Survivors (Fig. 2A). To provide multiplicity-controlled context, 12 of these nominal DEPs remained BH-FDR-supported (FSTL3, CKAP4, JUN, AGRN, LTBR, LAIR1, SPINK4, COLEC12, PRELP, IL7, TNFRSF11B, and PTH1R; Supplementary Table S4). Eleven were higher in the Death group, whereas IL7 was higher in Survivors. These proteins were used as a robustness reference rather than to redefine the exploratory DEP set. Proteins nominally upregulated in the Death group included FSTL3, CKAP4, AGRN, and LAIR1, whereas proteins higher in the Survivor group included CCL26, IL7, TNFSF10, and FASLG. This analysis summarized an outcome-associated proteomic contrast, defined by a set of mortality-associated proteins (e.g., FSTL3 and LAIR1) and a set of Survivor-associated proteins (e.g., CCL26), providing an exploratory candidate protein context.

Fig. 2.

Baseline serum proteomic signatures associated with 1-year survival status. A Volcano plot from limma analysis comparing Death and Survivor groups. The horizontal dashed line indicates P = 0.05. Red and blue dots indicate proteins meeting the nominal DEP definition of P < 0.05, with red dots upregulated in the Death group and blue dots higher in the Survivor group. B PCA biplot based on differentially expressed proteins (DEPs). Red diamonds indicate Death samples; blue triangles indicate Survivor samples. Arrows represent protein loading vectors. C Heatmap of top DEPs across samples. Columns are grouped by outcome and then clustered; rows are grouped by expression direction and then clustered. The color scale represents z-score normalized expression. Right-side bars indicate single-protein performance metrics, including AUC and VIP

A principal component analysis (PCA) biplot based on these DEPs showed partial separation between Death and Survivor samples (Fig. 2B). Loading vectors indicated that LAIR1, LTBR, and FSTL3 contributed to the Death-enriched, mortality-associated direction. The heatmap further supported outcome-related clustering of these DEPs across samples (Fig. 2C). Although individual proteins such as LAIR1 and FSTL3 showed single-marker discriminatory performance, with an area under the receiver operating characteristic curve (AUC) > 0.75, single-protein summaries did not fully capture the group-level proteomic pattern. This observation motivated the subsequent annotation-supported exploratory analysis.

Out-of-fold clinical-risk framework links proteomic signals to clinical risk

Having identified baseline proteomic signatures, we sought to relate these high-dimensional signals to an interpretable clinical-risk phenotype. The modeling workflow separated the pre-/post-EVT variable pool, the 246 model-eligible clinical variables after exclusion of seven leakage variables (Supplementary Table S5), the clinical-only elastic-net (ENet) out-of-fold (OOF) teacher model, and the resulting ClinIndex High/Low strata (Fig. 3A; detailed ClinIndex–ClinDistill workflow in Fig. S3A-B). It then linked these clinical-risk strata to the proteomic layer, including 345 pre-EVT serum proteins, 61 mortality-associated DEPs, weighted gene co-expression network analysis (WGCNA) mapping, the ClinDistill student representation, and the final 18 core proteins. In this workflow, the clinical-only OOF model generated mortality probabilities and defined the exploratory ClinIndex strata, whereas ClinDistill learned this clinical OOF risk signal in protein space and was not used to define ClinIndex. The clinical-only model achieved an internal OOF AUC of 0.825, with a Brier score of 0.141 (Fig. 3B). ClinDistill achieved a similar internal OOF AUC of 0.824, but was retained as an interpretability-oriented protein-space representation rather than as evidence of model superiority or external validation.

The clinical-only coefficient profile was driven by clinical factors such as Post-BUN, Post-NSE, Post-CRP, Age, and Pre-BUN, whereas Post-eGFR and Pre-GCS showed negative mortality-associated coefficients (Fig. 3C). The 18 non-zero ClinDistill proteins were carried forward as core proteins and subdivided by differential-expression status. Nine proteins that were also differentially expressed between Death and Survivor groups were defined as Core-DEP proteins (TNFRSF11B, FSTL3, CD276, CKAP4, CRIM1, LAIR1, IL4R, PRELP, and FABP1), whereas the remaining nine ClinDistill-selected proteins that were not differentially expressed were defined as Core-non-DEP proteins (SPON1, EPO, ESM1, CD160, CSF1, CEACAM21, VEGFD, SHMT1, and TIMP3). Five of the nine Core-DEP proteins (TNFRSF11B, FSTL3, CKAP4, LAIR1, and PRELP) were also BH-FDR-supported, linking part of the clinical-proteomic core signal to the multiplicity-controlled DEP subset. We next asked whether this risk-associated protein signal represented only a collection of individual proteins or a more organized biological association pattern. The cross-layer heatmaps summarized associations among selected clinical predictors, ClinDistill-derived core proteins, and WGCNA module eigengenes (Fig. 3D). Top clinical risk factors and top protein features shared module-level association patterns, supporting a descriptive clinical-protein-WGCNA module association structure rather than causal mediation.

Finally, we investigated the network context of the core subgroups by visualizing PPI annotation networks (Fig. 3E). The Core-DEP-centered network showed broader connectivity, with Core-DEP nodes such as CD276, CRIM1, PRELP, and FSTL3 connecting to immune and inflammatory proteins, including LILRB4 and TNFR/TNF-family-related nodes. In contrast, the Core-non-DEP-centered network was sparser and organized mainly around SPON1, EPO, ESM1, and TIMP3, with connections to adhesion/ECM- and tissue-structure-related proteins. This annotation-supported contrast suggested that Core-DEP and Core-non-DEP proteins occupied distinct biological contexts, motivating the deeper functional annotation presented in the next section.

Exploratory functional annotation summarizes candidate organization of selected protein groups

Building on this network contrast, we next performed a multiscale functional annotation to examine how the selected protein groups were organized. The DDI annotation workflow traced the analysis from the full measured protein panel through network filtering, retained interaction-supported nodes, DDI network construction, and annotation outputs (Fig. 4A). Fewer proteins appeared as network nodes after PPI/DDI filtering because only proteins with available annotation-supported interactions were retained in the final DDI network (Fig. 4B). At the domain-class level, Core-DEP-linked records showed a broad repertoire led by Immunoglobulin (446), Adhesion/extracellular matrix (Adhesion/ECM; 384), tumor necrosis factor receptor (TNFR; 194), Protease/Coagulation (147), leucine-rich repeat (124), and Other (109) classes. In contrast, Core-non-DEP-linked records were more concentrated and characterized by Adhesion/ECM (423), followed by Other (72), Immunoglobulin (39), Protease/Coagulation (28), leucine-rich repeat (25), and G protein-coupled receptor (5) classes (Fig. 4C).

At the functional-class interaction level, DDI chord diagrams further summarized this contrast. Core-DEP-linked records showed broader class-to-class connectivity involving immunoglobulin-, adhesion/ECM-, TNFR-related, cytokine/growth-factor-, and protease/coagulation-related classes, whereas Core-non-DEP-linked records were largely centered on Adhesion/ECM-related connectivity, consistent with a more focused functional organization (Fig. S4A-B).

At the domain-pair level, top DDI pairs provided a more granular view of this group-specific organization (Fig. 4D). Core-DEP-linked records featured immune-receptor and TNFR/TNF-related pairings, including immunoglobulin/TNF-family and TNF-family/TNFR-NGFR interactions, whereas Core-non-DEP-linked records were characterized by ECM, growth-factor, and tissue-structure-related pairings. Together, these results summarize a broader annotation-supported immune-receptor and TNFR/TNF-related candidate repertoire for Core-DEP proteins and a more focused Adhesion/ECM- and tissue-structure-related domain context for Core-non-DEP proteins.

ClinIndex maps to a candidate contrast between inflammatory and homeostatic patterns

To connect the functional vocabularies defined above with clinical risk, we prioritized ClinIndex-linked Core-Other candidate paths using the same PPI/DDI annotation framework. The path-scoring workflow integrated the ClinIndex Low/High split, protein-level Cohen’s d contrasts, W1-W2-W3 segment weights, bottleneck path ranking, function-pair aggregation, and domain-level interpretation (Fig. 5A). Inter-group PPI/DDI analysis identified annotation-supported bridge proteins across Core-DEP, Core-non-DEP, and Other-DEP sets, including core-side connectors such as CD276, CRIM1, LAIR1, TNFRSF11B, and SPON1, together with external bridge partners such as CXADR, LILRB4, CD4, CD40, and TNFR/TNF-family-related nodes (Fig. S4C). Among these, LAIR1 and TNFRSF11B were also BH-FDR-supported Core-DEP proteins. The bridge-protein heatmap summarized their DDI-count profiles across the three protein groups, supporting their interpretation as candidate connectors rather than causal drivers.

We next integrated the clinical-risk framework with domain-level interaction analysis in a four-ring summary view (Fig. 5B). Higher-risk (ClinIndex High, n = 54) and lower-risk (ClinIndex Low, n = 94) groups aligned with Core-DEP-centered and Core-non-DEP-centered candidate paths, respectively. Outcome sectors were included as contextual annotations and were not part of the path-scoring procedure. Because ClinIndex was derived from early clinical variables and serum proteins were measured before EVT, this analysis associated early clinical-risk phenotypes with baseline proteomic organization. Quantitative path aggregation showed that the higher-risk sector was characterized by high-weight immunoglobulin-tumor necrosis factor/tumor necrosis factor superfamily (TNF/TNFSF) and growth-factor-TNF/TNFSF class pairs, consistent with an annotation-supported TNFR/TNF-family candidate immune-interaction pattern. In contrast, the lower-risk sector showed stronger contributions from Other-Other, Immunoglobulin-Other, Lectin-Other, and growth-factor-related class pairs, consistent with an annotation-supported tissue-interaction and adhesion/ECM-oriented candidate pattern (Fig. 5C).

Finally, protein-domain-domain-protein (P-D-D-P) mapping translated these group-level vocabularies into representative protein-domain candidate chains (Fig. 5D). On the lower-risk side, candidate connections such as SPON1-COLEC12, SPON1-AGER, ESM1-CRELD2, EPO-PLAUR, and TIMP3-CXADR summarized an adhesion/ECM-oriented candidate pattern. On the higher-risk side, LAIR1-TNFSF11, TNFRSF11B-PDGFB, CRIM1-TNFSF10, and CD276-CD6 summarized a TNFR/TNF-family and immunoglobulin-domain candidate pattern. Thus, the ClinIndex-linked path analysis refined the broader Core-DEP/Core-non-DEP contrast into prioritized candidate connections, highlighting internally derived higher-risk inflammatory and lower-risk adhesion/ECM candidate patterns. Full PPI/DDI annotation records, Fig. 5 path-scoring records, and displayed P-D-D-P source records are provided in Supplementary Tables S6, S7, and S8, respectively.

Discussion

In this prospective EVT-treated ELVO cohort, we observed a 20.3% 1-year mortality rate despite a high reperfusion success rate. Early clinical indicators, including post-EVT dynamic variables, showed strong internal associations with mortality, and the clinical-only model achieved an internal OOF AUC of 0.825. Using the resulting ClinIndex as an exploratory stratification tool, we observed that 1-year prognosis was associated with two internally derived candidate baseline proteomic patterns: ClinIndex High status aligned with a TNFR/TNF-family-centered inflammatory pattern involving Core-DEP proteins, whereas ClinIndex Low status aligned with a more homeostatic Adhesion/ECM-centered pattern involving Core-non-DEP proteins.

The lower-risk pattern highlighted the potential importance of the Core-non-DEP subgroup, extending a DEP-centered interpretation by suggesting that clinically relevant biology may also be organized by proteins that do not individually meet the DEP threshold. Favorable outcomes may reflect preservation of a homeostatic structural framework [29]. Consistent with this interpretation, the Core-non-DEP group was characterized by Adhesion/ECM functions and thrombospondin/ECM-related domain annotations. In the P-D-D-P analysis, representative lower-risk candidate connections included SPON1-COLEC12, SPON1-AGER, ESM1-CRELD2, EPO-PLAUR, and TIMP3-CXADR, supported by thrombospondin type 1, erythropoietin/thrombopoietin, insulin-like growth factor binding protein, lectin C-type, u-PAR/Ly-6, and tissue inhibitor-related domain pairings [30, 31]. These annotation-supported links define candidate structural hypotheses for future validation. CXADR is a tight-junction component essential for blood-brain barrier integrity [32], supporting the biological plausibility of an adhesion/ECM-related pattern associated with favorable outcomes. Thus, this pattern may reflect preserved neurovascular-unit integrity and resistance to inflammatory degradation [33].

By contrast, the higher-risk pattern was linked to the TNFR/TNF-family axis [34, 35], with Core-DEP-linked records showing a broad immune/receptor-related domain profile that included Immunoglobulin, TNFR, Protease/Coagulation, and cytokine/growth-factor ligand classes [36]. Within this annotation-supported network, representative mortality-associated candidate connections included LAIR1-TNFSF11, TNFRSF11B-PDGFB, CRIM1-TNFSF10, and CD276-CD6, supported by immunoglobulin, TNFR/NGFR cysteine-rich, TNF-family, von Willebrand factor type C, scavenger receptor cysteine-rich, and PDGF/VEGF domain pairings. These findings support a TNFR/TNF-family-centered candidate inflammatory pattern enriched for TNFR/TNF-family and immunoglobulin-domain candidate interaction patterns. TNFRSF11B has also been linked to ischemic stroke risk in prior genetic association work [37]. These annotation-supported priorities nominate candidate inflammatory patterns for future validation [38, 39].

A core methodological feature of this study is its clinically anchored but biologically interpretable framework. The clinical-only model defined the ClinIndex strata, whereas ClinDistill represented the clinical OOF risk signal in protein space. Although ClinDistill achieved a similar internal OOF AUC of 0.824, it was retained for interpretation rather than as evidence of model superiority or external validation. This exploratory approach links clinical risk factors, such as Age and Post-BUN [40], to candidate inflammatory patterns while preserving biological interpretability beyond feature-importance summaries [19, 41]. This decoupling also helps interpret the SPON1 signal: although SPON1 contributed strongly to the ClinDistill risk representation, annotation-based analysis placed SPON1 within lower-risk thrombospondin-domain candidate connections such as SPON1-COLEC12, SPON1-AGER, and SPON1-FST. One plausible interpretation is that SPON1 is context-dependent, and its cohort-wide statistical association with mortality may reflect disruption of a homeostatic pattern rather than a dominant causal role.

Because serum samples were collected before EVT, the proteomic findings represent baseline inflammatory contexts associated with subsequent early clinical-risk status and 1-year mortality, rather than molecular consequences of EVT itself. Several interpretive boundaries guide the translational reading of this framework. ClinIndex was used as an early post-EVT clinical-risk phenotype to organize baseline proteomic patterns, and the Youden-derived ClinIndex threshold supported exploratory stratification and visualization. Similarly, the PPI, DDI, and P-D-D-P analyses provide annotation-based prioritization of candidate protein-domain patterns rather than direct mechanistic proof. Therefore, the TNFR/TNF-family inflammatory pattern should be interpreted as a baseline inflammatory state associated with higher clinical risk, not as a validated causal pathway [42, 43].

This study has limitations. First, the single-center cohort size and modest number of death events (n = 148, 30 deaths) may limit coefficient stability, and multicenter validation is required. Second, although all-cause 1-year mortality was prespecified as the primary endpoint, it may underrepresent disability burden; in this cohort, unfavorable 1-year functional outcome (mRS 4–6) occurred in 41.2% of patients (61/148), supporting future endpoint frameworks that jointly model mortality and disability. Third, detailed cause-of-death and timing analyses should be expanded to refine stroke-specific biological interpretation. Fourth, because the DEP screen involved multiple protein tests and was designed as an exploratory screening analysis rather than a multiplicity-adjusted confirmatory discovery analysis, DEP-based findings should be interpreted as exploratory adjusted baseline protein signals rather than confirmed findings. Fifth, interaction findings were derived from STRING/PPI-ID annotations and require direct functional validation [44]. Finally, the proteomic framework was based on pre-EVT samples only, and longitudinal post-EVT proteomic trajectories warrant a dedicated study. In addition, selected protein signals and interaction annotations may reflect cohort-specific patterns rather than reproducible biological signals. The candidate protein patterns identified here should therefore be viewed as hypotheses that require independent EVT cohorts and functional studies before clinical or mechanistic interpretation.

In conclusion, this study used an interpretable clinical-proteomic framework to examine baseline proteomic patterns associated with 1-year mortality after EVT. We observed two internally derived candidate patterns linked to early post-EVT clinical-risk phenotypes: a TNFR/TNF-family-centered inflammatory pattern involving Core-DEP proteins and a more homeostatic Adhesion/ECM-centered pattern involving Core-non-DEP proteins. The higher-risk candidate pattern was represented by candidate connections such as LAIR1-TNFSF11 and TNFRSF11B-PDGFB, whereas the lower-risk candidate pattern was represented by SPON1-COLEC12, SPON1-AGER, and TIMP3-CXADR. These findings provide a neuroinflammation-focused, hypothesis-generating framework for understanding proteomic patterns associated with poor outcomes despite reperfusion and nominate candidate patterns for validation in independent cohorts and disease models.

Authors’ contributions

Conception and design: Jixian Lin, Min Xiao, Niu Qiao. Collection and assembly of data: Jixian Lin, Ruihong Zhang, Rong Xi, Xiaojun Ma, Yang Liu, Huanyin Li. Data analysis and interpretation: Jixian Lin, Min Xiao, Niu Qiao. Manuscript writing: Jixian Lin, Min Xiao, Niu Qiao, Hai-Qian Zhou. Final approval of manuscript: All authors. Accountable for all aspects of the work: All authors.

Funding

This work is supported by Shanghai Municipal Health Commission Medical Special Project (202240231), Medical Engineering Joint Fund of Fudan University (yg2022-32), Shanghai Natural Science Foundation Project (25ZR1402038).

Data availability

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

Declarations

Ethics approval and consent to participate

The study was registered in the Chinese Clinical Trial Registry (Registration number: ChiCTR2100045753). Ethical approval was obtained from the institutional review board at Minhang Hospital, Fudan University, Shanghai, China. Written informed consent was obtained from all participants or their legally authorized representatives before enrollment.

Competing interests

The authors declare no competing interests.