Untargeted red blood cell metabolomics reveals distinct metabolic signatures in malignant cerebral edema following acute ischemic stroke

Abstract

Background

Malignant cerebral edema (MCE) is a major cause of mortality and disability in acute ischemic stroke, yet effective treatments remain limited. Early identification of patients at high risk for MCE is therefore an important unmet clinical need.

Methods

Patients with large vessel occlusion acute ischemic stroke (LVO-AIS) were prospectively enrolled. All patients had radiologically confirmed LVO and underwent follow-up neuroimaging for cerebral edema assessment. MCE was defined as cerebral edema grade 3 (CED-3) on the Thrombolysis in Stroke–Monitoring Study edema scale, while non-MCE was defined as CED-1–2. In a pilot screening phase, plasma, red blood cell (RBC), and urine samples were collected on days 1, 3, and 7 after stroke onset from a subset of patients to identify the most informative biofluid and sampling time point. Based on these results, untargeted metabolomic profiling using ultrahigh-performance liquid chromatography–tandem mass spectrometry was performed in RBC samples collected within 24 h of stroke onset from an expanded cohort.

Results

Pilot screening (n = 3 per group) indicated that day 1 RBC samples showed the most pronounced metabolic differences between patients who later developed MCE and those who did not. Accordingly, day 1 RBCs from 61 patients (19 MCE, 42 non-MCE) were analyzed. Untargeted metabolomic analysis identified 521 differentially abundant metabolites, including 177 upregulated and 344 downregulated features in the MCE group. Key altered metabolites included palmitoylethanolamide, S-adenosylmethionine, and spermidine, with enrichment in pathways related to cofactor biosynthesis, neuroactive ligand–receptor interactions, and amino acid metabolism.

Conclusion

Red blood cell metabolomic profiles within 24 h of stroke onset differ between patients who subsequently develop malignant cerebral edema and those who do not. These findings provide exploratory insights into early metabolic alterations associated with MCE and warrant further validation in larger, targeted studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12883-026-04684-0.

Introduction

Large vessel occlusion (LVO) is present in approximately 15–30% of patients with acute ischemic stroke (AIS) [1]. Approximately one-quarter of LVO patients present with large territorial infarctions at admission [2], and more than 50% of these patients may develop malignant cerebral edema (MCE), a condition characterized by severe brain swelling, midline shift, and herniation, which substantially increases the risk of death [3]. The mortality rate can reach 80%, and the only proven life-saving intervention is decompressive hemicraniectomy performed within 48 hours [4]. However, only 0.3% of stroke patients are eligible for this surgical procedure. Timely identification of patients at high risk for MCE is therefore a major unmet clinical need.

Timely identification of patients at high risk for MCE is therefore a major unmet clinical need, as MCE represents a time-sensitive and potentially actionable condition in which early biological signals may directly influence monitoring intensity and therapeutic decision-making. Although effective pharmacological therapies for MCE are currently lacking, increasing evidence suggests that a deeper understanding of its underlying pathophysiological mechanisms and predictive factors may substantially improve early risk stratification.

Previous studies have reported multiple circulating biomarkers associated with MCE development, including metabolic markers (e.g., blood glucose) [5], inflammatory indices (e.g., neutrophil-to-lymphocyte ratio, neutrophil count, procalcitonin) [6, 7], endothelial and blood–brain barrier–related proteins (e.g., S100B, cellular fibronectin) [8, 9], lipid mediators such as 20-hydroxyeicosatetraenoic acid [10, 11], and neurohumoral factors. However, most of these studies were limited by small sample sizes, retrospective designs, and a focus on single biomarkers.

In parallel, several prediction models for MCE have been proposed. Nevertheless, their clinical applicability remains limited due to highly selective inclusion criteria, retrospective data collection, modest discriminative performance, lack of external validation, and small cohorts [11]. Importantly, both existing biomarkers and prediction models largely fail to capture the complex biological mechanisms driving cerebral edema formation, resulting in limited specificity for MCE prediction.

Therefore, there remains a critical need for prospective studies that adopt mechanism-oriented approaches to identify more specific biomarkers capable of capturing the multifaceted metabolic disturbances associated with malignant cerebral edema.

The development of MCE involves complex pathophysiological processes, including metabolic dysregulation, inflammation, blood–brain barrier disruption, and oxidative stress [12–14]. Untargeted metabolomics is a powerful approach in metabolomic research, enabling comprehensive profiling of metabolites in biological systems through high-throughput analysis without prior knowledge of the target compounds [15]. This technique simultaneously captures multiple metabolic pathways, facilitating the characterization of metabolic signatures associated with disease onset and progression [16, 17]. Metabolites play crucial roles in physiological and pathological processes and are measurable in diverse biofluids, including plasma, urine, and red blood cells (RBCs). Their composition is influenced by both internal and external environmental factors, and disease states may disrupt specific metabolic pathways, making them potential biomarkers [18, 19]. Notably, certain classes of metabolites, including amino acids, acylcarnitines, sphingolipids, and other small polar or lipid-soluble molecules, are able to cross the blood–brain barrier (BBB) through passive diffusion or specific transport systems. Consequently, alterations in their levels in peripheral blood may partially reflect metabolic disturbances within the central nervous system. Previous metabolomics studies have demonstrated that circulating metabolites such as branched-chain amino acids, acylcarnitines, sphingolipids, and tricarboxylic acid cycle intermediates are associated with stroke presence, infarct burden, and stroke aetiology, supporting the biological plausibility of blood-based metabolomic markers in cerebrovascular disease [18, 19]. These findings provide novel insights into the pathophysiological mechanisms of MCE following AIS [20].

This study utilized untargeted metabolomics based on ultrahigh-performance liquid chromatography–tandem mass spectrometry (UPLC-MS/MS) to examine plasma, RBC, and urine samples from patients with LVO. The metabolic profiles of the MCE and non-MCE groups were compared to identify metabolic alterations associated with MCE and to elucidate the underlying pathophysiological mechanisms, with the aim of enhancing MCE prediction and facilitating targeted clinical interventions.

Methods

Study design

This study consisted of two sequential phases.

Phase I was an exploratory pilot screening study designed to evaluate the feasibility of multi-biofluid untargeted metabolomics and to determine the optimal biofluid and sampling time point associated with MCE. Six patients with LVO-AIS, including three patients who later developed MCE and three who did not, were selected based on comparable baseline characteristics. Plasma, RBC, and urine samples were collected within 24 h, on day 3, and on day 7 after stroke onset.

Phase II was the main analysis. Based on the findings from Phase I, untargeted metabolomics was subsequently performed exclusively on day-1 RBC samples from an expanded cohort. All samples in this phase were collected within 24 h after stroke onset, and patients were retrospectively classified into MCE and non-MCE groups according to follow-up imaging performed between days 3–7 after stroke onset.

Study population

Patients with LVO admitted to the participating center between November 2023 and December 2024 were screened. Eligible patients were enrolled when biospecimens could be obtained within 24 h after stroke onset, according to the predefined inclusion and exclusion criteria.

According to the Trial of Org 10,172 in Acute Stroke Treatment (TOAST) classification, large vessel occlusion in acute ischemic stroke predominantly arises from large-artery atherosclerosis (LAA) and cardioembolism (CE). Consistent with this epidemiological pattern, the majority of patients included in the present cohort were classified as LAA or CE, while other etiological subtypes were rare.

The inclusion criteria were as follows: (1) aged ≥ 18 years; (2) hospital admission within 24 h of stroke onset; and (3) diagnosed in accordance with the 2018 Chinese Guidelines for Acute Ischemic Stroke Management. (4) LVO was confirmed by neuroimaging modalities, identifying occlusions in arteries such as the distal internal carotid artery, carotid T junction, A1 segment of the anterior cerebral artery, M1 segment of the middle cerebral artery, V4 segment of the vertebral artery, basilar artery, or P1 segment of the posterior cerebral artery. Imaging techniques included computed tomography angiography (CTA), computed tomography perfusion (CTP), magnetic resonance angiography (MRA), and digital subtraction angiography (DSA). (5) Serum albumin level > 30 g/L. (6) Informed consent was obtained from patients or their legal guardians. The exclusion criteria were as follows: (1) loss to follow-up or missing key clinical, imaging, or laboratory data required for patient classification or primary analysis, such as admission NIHSS score, cerebral edema grading based on follow-up CT/MRI, or failure of biospecimen collection; (2) having a coexisting malignancy or hematological disorders; (3) having bleeding disorders, immunosuppression, or active infections; (4) having severe cardiac, hepatic, or renal dysfunction; and (5) having a prestroke modified Rankin scale (mRS) score > 2.

Importantly, none of the enrolled patients met criteria for MCE at admission. All patients classified as MCE developed cerebral edema during hospitalization and were categorized based on the most severe cerebral edema observed on follow-up CT or MRI performed between days 3 and 7 after stroke onset.

Clinical data collection

We recorded demographic data (age, sex), medical and personal history, and admission clinical scores, including the National Institutes of Health Stroke Scale (NIHSS) [21], Glasgow Coma Scale (GCS) [22]. The mRS score [23] referred to premorbid functional status prior to stroke onset and was used as an exclusion criterion, with patients having a prestroke mRS > 2 excluded from the study. The laboratory test results included fasting glucose, D-dimer, glycated hemoglobin, and other relevant biochemical indices.

Admission and follow-up CT or MRI data ( between days 3 and 7 after stroke onset ) were assessed for cerebral edema severity (CED), which was graded via the Thrombolysis in Stroke-Monitoring Study (SITS-MOST) edema scale [24]. Cerebral edema grades were defined as follows: CED-1 for involvement of ≤ 1/3 hemisphere, CED-2 for > 1/3 hemisphere, and CED-3 for edema accompanied by midline shift. The day with the most severe edema was selected for classification. Patients with CED-3 were classified into the MCE group, and all others were assigned to the non-MCE group [25].

Sample preparation

Fasting venous blood and urine samples were collected on days 1, 3, and 7 after stroke onset according to standardized biospecimen collection procedures. Blood samples were collected into ethylenediamine tetraacetic acid (EDTA) anticoagulant tubes (Bioscience Tian’ai, China). Red blood cells were isolated by centrifugation, followed by careful removal of plasma and the buffy coat. The packed erythrocytes were washed three times with isotonic HEPES-buffered saline (pH 7.4) to remove residual plasma components. All samples were processed under light-protected conditions and stored in ultralow-temperature freezers (− 80 °C) until analysis. Repeated freeze–thaw cycles were strictly avoided.

For metabolomic analysis, plasma, RBC, and urine samples were processed according to standardized protocols. Briefly, 100 µL of plasma or urine, or 25 mg of RBC pellets, were mixed with precooled extraction solvent containing isotopically labeled internal standards (d3-leucine, 13C9-phenylalanine, d5-tryptophan, and 13C3-progesterone) (methanol: acetonitrile: water = 4:2:1, v/v/v for plasma and urine; 2:2:1, v/v/v for RBCs). The mixtures were vortexed or homogenized and incubated at − 20 °C for 2 h, followed by centrifugation at 25,000 × g for 15 min at 4 °C.

A total of 600 µL of the supernatant was transferred and dried under vacuum. The dried extracts were reconstituted in aqueous organic solvent prior to UPLC–MS/MS analysis: plasma and urine extracts were reconstituted in 180 µL of methanol–water (1:1, v/v), while RBC extracts were reconstituted in 120 µL of 50% methanol. After centrifugation at 25,000 × g for 15 min at 4 °C, the supernatants were collected for analysis. Quality control (QC) samples were prepared by pooling equal aliquots from each individual sample.

UPLC‒MS/MS analysis

Metabolite separation was performed using a UPLC I-Class Plus system (Waters, USA) equipped with a BEH C18 column (1.7 μm, 2.1 × 100 mm; Waters, USA). The column temperature was maintained at 45 °C, and the flow rate was set to 0.35 mL/min. The injection volume was 5 µL.

For positive ion mode, the mobile phase consisted of water containing 0.1% formic acid (A) and methanol containing 0.1% formic acid (B). For negative ion mode, the mobile phase consisted of water containing 10 mM ammonium formate (A) and 95% methanol containing 10 mM ammonium formate (B).

The gradient elution program was as follows: 0–1 min, 2% B; 1–9 min, linear increase from 2% to 98% B; 9–12 min, maintained at 98% B; 12–12.1 min, rapidly decreased from 98% to 2% B; and 12.1–15 min, re-equilibration at 2% B.

Mass spectrometric detection was performed on a Q Exactive high-resolution mass spectrometer (Thermo Fisher Scientific, USA) operated in both positive and negative electrospray ionization modes. Full MS scans were acquired over an m/z range of 70–1050 at a resolution of 120,000, with an automatic gain control (AGC) target of 3 × 10⁶ and a maximum injection time of 100 ms. Data-dependent MS/MS (ddMS²) spectra were obtained for the top three most intense precursor ions at a resolution of 30,000, with an AGC target of 1 × 10⁵, a maximum injection time of 50 ms, and stepped normalized collision energies of 20%, 40%, and 60%.

The spray voltage was set to + 3.80 kV in positive mode and − 3.20 kV in negative mode. The capillary temperature was maintained at 320 °C, and the auxiliary gas heater temperature was set to 350 °C. Sheath gas and auxiliary gas flow rates were set to 40 and 10, respectively.

Metabolomic data processing

Data processing was conducted using Compound Discoverer software (version 3.3, Thermo Fisher Scientific) for peak detection, retention time alignment, feature grouping, and metabolite annotation. Detected signals were initially defined as LC–MS features characterized by unique m/z–retention time pairs.

Putative metabolite annotation was performed by matching accurate precursor masses, isotope patterns, and, when available, MS/MS spectra against public databases including the Human Metabolome Database (HMDB), mzCloud, and ChemSpider. A mass accuracy tolerance of ≤ 5 ppm was applied for precursor ion matching. MS/MS spectral similarity matching was prioritized when high-quality fragmentation spectra were available; otherwise, annotation was based on accurate mass and isotope pattern.

Common adduct forms were considered during annotation, and isotopic peaks and in-source fragments were filtered using the default Compound Discoverer workflows. According to the Metabolomics Standards Initiative (MSI) criteria, most annotated metabolites correspond to MSI Level 2 (putatively annotated compounds based on MS/MS similarity) or Level 3 (putatively characterized compound classes). No Level 1 identifications using authentic standards were claimed in this untargeted analysis.

Probabilistic quotient normalization (PQN) and QC-based robust LOESS signal correction (QC-RLSC) were applied. Features exhibiting a coefficient of variation (CV) > 30% in quality control samples were excluded. Multivariate analyses including PCA, PLS-DA, and OPLS-DA were used for exploratory visualization of group-wise metabolic patterns. Data quality and system stability were monitored using PCA and base peak chromatogram (BPC) overlays [26, 27].

Statistical analysis

Given the exploratory nature of untargeted metabolomics, no formal a priori sample size calculation was performed. The sample size was determined by the number of eligible patients with complete biospecimen availability during the study period.

Multivariate analyses, including PCA and OPLS-DA, were used for pattern recognition and visualization of metabolic differences between groups rather than for primary feature selection. Differential feature selection was performed using a combined strategy incorporating multivariate analysis and univariate statistics. Features with a variable importance in projection (VIP) score ≥ 1 from PLS-DA/OPLS-DA, an absolute fold change ≥ 1.2 or ≤ 0.83, and a nominal p value < 0.05 were considered differentially abundant.

Given the exploratory nature of this untargeted metabolomics study, univariate p values were not adjusted for multiple testing and were used as an initial screening criterion to prioritize features for downstream exploratory analysis rather than for definitive statistical inference.

Volcano plots were generated using GraphPad Prism (v10.1.2), and receiver operating characteristic (ROC) curve analysis was used to evaluate the discriminative performance of selected features (AUC > 0.70).

Statistical analysis of baseline clinical data was performed using SPSS version 27.0. Depending on data distribution, the Kolmogorov–Smirnov test, independent samples t test, Mann–Whitney U test, or chi-square test was applied. A two-sided p value < 0.05 was considered statistically significant.

Results

Study flow and sample selection

The overall study flow and sample selection process are summarized in Fig. 1. Briefly, this study was conducted in two sequential phases. In Phase I (pilot screening), a small subset of patients was analyzed across multiple biofluids and time points to identify the most informative matrix and sampling window for malignant cerebral edema–related metabolic alterations. In Phase II, based on the pilot results, an expanded cohort was included for untargeted metabolomic analysis focusing on day-1 red blood cell samples. Details regarding patient screening, exclusion, sample availability, and group allocation are illustrated in the study flowchart.

Fig. 1.

Study design and patient flowchart

Pilot experiment to determine the optimal biofluid and sampling time point

Untargeted metabolomic profiling was performed as a pilot screening on plasma, RBC, and urine samples collected on days 1, 3, and 7 after stroke onset from three patients with MCE and three with non-MCE (Additional file 1). In the RBC samples, 12,687 metabolic features were detected, with the hightest number of differentially abundant features observed on day 1 (n = 518). In the plasma samples, 5,610 features were identified, of which 118 showed significant different on day 1. In urine samples, 21,850 features were detected, with the largest number of differential features observed on day 3 (n = 544), followed by day 1 (n = 263).

Overall, this pilot screening suggested that, compared with plasma and urine, the RBC metabolome exhibited greater sensitivity to early metabolic differences between the MCE and non-MCE groups. Notably, distinct metabolic alterations were detectable as early as day 1 after stroke onset, supporting the selection of day-1 RBC samples for subsequent analyses in the expanded cohort.

Baseline characteristics of the main study cohort

Based on the results of the pilot screening phase, red blood cell samples collected within 24 h of stroke onset were analyzed in an expanded cohort comprising 61 patients with large vessel occlusion acute ischemic stroke, including 19 patients in the MCE group and 42 in the non-MCE group. Compared with non-MCE patients, MCE patients presented significantly greater systolic blood pressure (162.53 ± 32.64 vs. 146.53 ± 25.13, p = 0.035), higher NIHSS scores at admission (13 [12, 18] vs. 8 [5, 15], p = 0.003), higher 24-hour NIHSS scores (17 [13, 20] vs. 7 [3, 12], p < 0.001), lower NIHSS differences (0 [-6, 0] vs. 2 [0, 5], p = 0.020), and lower lymphocyte counts (0.41 [0.29, 0.71] vs. 1.16 [0.91, 1.91], p = 0.026) than non-MCE patients did (Table 1).

Table 1.

Baseline data

Variables	MCE(n = 19)	non-MCE(n = 42)	t/χ2/Z	P value
Demographics
Age, y; median (IQR)	77 (66,82)	71 (59,81)	-1.091#	0.275
Male; (n%)	7 (11.5%)	20 (32.8%)	0.616¶	0.579
Previous History n (%)
Hypertension; n (%)	9 (14.8%)	22 (36.1%)	0.132¶	0.786
DM; n (%)	4 (6.6%)	15 (24.6%)	1.311¶	0.376
CAD; n (%)	5 (8.2%)	5 (8.2%)	1.982¶	0.261
HL; n (%)	5 (8.2%)	18 (29.5%)	1.524¶	0.264
AF; n (%)	11 (18.0%)	19 (31.1%)	0.838¶	0.416
Stroke; n (%)	0 (0.0%)	5 (8.2%)	2.464¶	0.313
SD; (n%)	0 (0.0%)	5 (8.2%)	2.464¶	0.313
Smoking; n (%)	4 (6.6%)	11 (18.0%)	0.186¶	0.757
Drinking; n (%)	2 (3.3%)	9 (14.8%)	1.052¶	0.476
Baseline Parameters
Temperature; median (IQR)	36.5 (36.2,36.7)	36.5 (36.5,36.8)	-1.074#	0.283
HR; mean ± SD	84.89 ± 17.92	82.48 ± 17.84	-0.490†	0.626
SBP (mmHg); mean ± SD	162.53 ± 32.64	146.02 ± 25.13	-2.160†	0.035
DBP (mmHg); median (IQR)	87 (65, 95)	79 (67, 87)	-1.699#	0.089
Initial NIHSS Score; median (IQR)	13 (12, 18)	8 (5, 15)	-3.011#	0.003
Initial GCS Score; median (IQR)	13 (5, 14)	14 (10, 15)	-1.480#	0.139
24 h NIHSS score; median (IQR)	17 (13, 20)	7 (3, 12)	-4.164#	< 0.001
NIHSS difference; median (IQR)	0 (-6, 0)	2 (0, 5)	-2.331#	0.020
Laboratory Data
Glu; median (IQR)	13.70 (11.95, 19.15)	16.70 (13.40, 20.43)	-1.247#	0.212
Neutrophil (×109/L); median (IQR)	7.58 (5.92, 11.31)	6.07 (4.37, 8.75)	-1.057#	0.178
Monocyte (×109/L); median (IQR)	0.41 (0.29, 0.71)	0.48 (0.36, 0.69)	-0.516#	0.606
Lymphocyte (×109/L); median (IQR)	0.41 (0.29, 0.71)	1.16 (0.91, 1.91)	-2.223#	0.026
HB (g/L); mean ± SD	125.50 ± 19.38	129.40 ± 23.08	0.628†	0.532
PLT (×109/L); median (IQR)	157 (115, 208)	174 (126, 219)	-0.653#	0.513
TT; median (IQR)	14.85 (13.75, 16.05)	14.30 (13.70, 15.90)	-0.277#	0.674
D-dimer (mg/L); median (IQR)	834.50(325.25,1748.50)	401.90(146.25,1093.50)	-1.614#	0.107
INR; mean ± SD	1.02 ± 0.11	1.04 ± 0.09	1.151†	0.255
FIB (g/L); median (IQR)	3.22 (2.51, 3.72)	3.13 (2.50, 3.55)	-0.454#	0.650
Cr (µmol/L); median (IQR)	67.70 (49.30, 86.80)	73.85 (60.75, 88.63)	-1.152#	0.249
BUN (µmol/L); median (IQR)	5.94 (4.82, 8.49)	5.86 (4.75 ,7.20)	-0.909†	0.606
UA (µmol/L); median (IQR)	347.89 ± 160.22	311.89 ± 135.21	-0.516#	0.367
Apolipoprotein B; mean ± SD	1.00 ± 0.33	0.86 ± 0.20	-1.650†	0.113
Lp (a); median (IQR)	234.60 (66.23, 402.58)	185.10 (108.25, 307.20)	-0.280#	0.780
ALB (g/L); median (IQR)	38.40 (35.68, 40.68)	38.20 (34.90, 41.35)	-0.161†	0.873
TC (mmol/L); median (IQR)	4.48 (3.86, 5.76)	4.42 (3.34, 5.27)	-0.420#	0.675
LDL (mmol/L); median (IQR)	2.55 (2.10, 3.44)	2.40 (1.80, 3.27)	-0.905#	0.365
HbA1c(%); median (IQR)	8.30 (2.50, 22.80)	4.30 (1.10, 10.00)	-1.667#	0.096
Hcy (µmol/L); median (IQR)	6.40 (5.90, 6.78)	6.15 (5.10, 7.45)	-0.703#	0.482

Abbreviations: MCE Malignant Cerebral Edema, DM Diabetes Mellitus, CAD Coronary Artery Disease, HL Hyperlipidemia, AF Atrial Fibrillation, SD Sleep Disorder, HR Heart Rate, SBP Systolic Blood Pressure, DBP Diastolic Blood Pressure, NIHSS National Institutes of Health Stroke Scale, mRS Modified Rankin Scale, GCS Glasgow Coma Scale, Glu Glucose, HB Hemoglobin, PLT Platelet, TT Thrombin time, INR International Normalized Ratio, FIB Fibrinogen, Cr Creatinine, BUN Blood urea nitrogen, UA Uric acid, Lp (a) Lipoprotein (a), ALB Albumin, TC Total cholesterol, LDL Low-density lipoprotein, HbA1c Hemoglobin A1c, Hcy Homocysteine

†, Student’s t test

#, Mann‒Whitney U test

¶, Chi‒square test

RBC metabolic profiles

A total of 32,254 LC–MS features were detected, of which 8,350 were annotated as putative metabolites based on accurate mass and MS/MS spectral matching. PCA, PLS-DA, and OPLS-DA were applied to explore overall metabolic patterns between the MCE and non-MCE groups, revealing partial group-wise clustering in the score plots (Fig. 2).

Fig. 2.

Multivariate analysis of differential metabolites. Notes: Part A Principal component analysis (PCA) score plot showing the overall distribution of samples. The percentages in parentheses indicate the variance explained by each principal component. Part B Partial least squares discriminant analysis (PLS-DA) score plot illustrating group-level metabolic pattern differences. Part C Orthogonal PLS-DA (OPLS-DA) score plot demonstrating partial group-wise clustering between the MCE and non-MCE groups. Part D Permutation test of the OPLS-DA model. The negative Q² intercept indicates limited predictive performance, supporting the use of OPLS-DA for exploratory pattern visualization rather than predictive modeling. The ellipse represents the 95% confidence interval

Model validation using permutation testing demonstrated a high explained variance (R2Y = 0.922) but a negative predictive variance (Q2 = − 0.0156), indicating limited predictive performance of the OPLS-DA model (Supplemental Table 2). Accordingly, multivariate models were used for exploratory pattern recognition and visualization rather than for predictive discrimination.

Differentially abundant metabolic features and annotated metabolites in the MCE and non-MCE groups

Differentially abundant LC–MS features were initially identified via the following thresholds: (1) VIP ≥ 1; (2) fold change ≥ 1.2 or ≤ 0.83; and (3) p < 0.05. Based on these thresholds, 521 differential features were identified, including 344 upregulated and 177 downregulated features between the MCE and non-MCE groups.

These differential features were subsequently annotated as putative metabolites based on accurate mass and MS/MS spectral matching against reference databases. From the annotated metabolite list, 17 representative metabolites were selected for focused downstream analysis based on a combination of statistical significance, effect size, annotation confidence, and biological relevance to ischemic brain injury (Table 2; Fig. 3).

Table 2.

Annotated metabolites selected from differentially abundant features

Metabolite Name	tR/min	Molecular Formula	VIP	P value	Fold Change	Status
spermine	0.588	C10H26N4	1.69	0.027	0.73	Down
phenylalanylphenylalanine	3.487	C12H16N2O3	2.78	0.033	1.93	Up
glycyl-L-leucine	1.791	C8H16N2O3	2.78	0.002	1.67	Up
D-aspartic acid	3.820	C4H7NO4	1.35	0.034	0.70	Down
DL-glutamine	3.954	C5H10N2O3	1.38	0.020	0.71	Down
L-γ-glutamylcysteine	1.066	C8H14N2O5S	2.09	0.037	1.54	Up
S-adenosylmethionine	6.025	C15H22N6O5S	1.82	0.023	0.64	Down
α-aspartylphenylalanine	3.790	C13H16N2O5	1.46	0.049	0.63	Down
butyrylcarnitine	2.890	C11H21NO4	2.01	0.040	1.73	Up
Palmitoyl ethanolamide	8.600	C18H37NO2	2.11	0.034	1.47	Up
Phosphatidylglycerol(18:1/0:0)	9.725	C24H47O9P	1.67	0.022	1.30	Up
Phosphatidylcholine(16:0/20:4)	11.017	C44H80NO8P	1.68	0.033	0.57	Down
2-methylglutaric acid	0.605	C6H10O4	1.68	0.001	0.71	Down
isobutyrylglycine	1.357	C6H11NO3	1.37	0.024	0.76	Down
dodecanoic acid	9.106	C12H24O2	1.69	0.029	0.77	Down
glutaric acid	2.983	C5H8O4	1.90	0.044	0.74	Down
3-Hydroxyhippuric acid	3.414	C9H9NO4	2.06	0.009	0.32	Down

Status indicates the direction of change of each metabolite in the MCE group relative to the non-MCE group (Up: increased; Down: decreased)

Fig. 3.

Volcano plot of RBC samples between the MCE and non-MCE groups. Notes: A total of 521 differentially abundant metabolites were identified, including 344 upregulated and 177 downregulated metabolites. Each dot represents an annotated metabolite. Upregulated metabolites are shown in orange, downregulated metabolites in blue, and non-significant metabolites in gray. Seventeen representative metabolites selected for focused analysis are highlighted in green and labeled. Abbreviations: C4-Carn, butyrylcarnitine; 2-MeGlu, 2-methylglutaric acid; 3-OH-Hip, 3-hydroxyhippuric acid; D-Asp, D-aspartic acid; DL-Gln, DL-glutamine; GluA, glutaric acid; Gly-Leu, glycyl-L-leucine; iBu-Gly, isobutyrylglycine; Phe-Phe, phenylalanylphenylalanine; PEA, palmitoylethanolamide; PC (16:0/20:4), phosphatidylcholine (16:0/20:4); PG (18:1/0:0), phosphatidylglycerol (18:1/0:0); SAM, S-adenosylmethionine; α-Asp-Phe, α-aspartylphenylalanine; C12:0, dodecanoic acid; SP, spermine

Among the annotated metabolites, the downregulated compounds included spermine, D-aspartic acid, α-aspartylphenylalanine, DL-glutamine, S-adenosylmethionine (SAM), phosphatidylcholine (PC, 16:0/20:4), 2-methylglutaric acid, isobutyrylglycine, dodecanoic acid, glutaric acid, and 3-hydroxyhippuric acid. The upregulated annotated metabolites included phenylalanylphenylalanine, glycyl-L-leucine, L-γ-glutamylcysteine, butyrylcarnitine, palmitoylethanolamide (PEA), and phosphatidylglycerol (PG, 18:1/0:0)

ROC analysis was performed as an exploratory assessment of the discriminatory potential of these 17 representative metabolites. Glycyl-L-leucine showed the highest area under the curve (AUC = 0.739), whereas the AUC values of the remaining metabolites were all below 0.70 (Fig. 4)

Fig. 4.

ROC curve of glycyl-L-leucine. Notes: ROC curves were generated to evaluate the discriminative performance of the 17 representative metabolites between the MCE and non-MCE groups. Glycyl-L-leucine (1.791_187.1088151 is the KEGG metabolite ID for glycyl-L-leucine, which uniquely identifies this metabolite in the KEGG database) showed the highest area under the curve (AUC = 0.739), whereas the remaining metabolites exhibited lower AUC values. These results indicate modest discriminative performance and are presented for exploratory purposes

Pathway enrichment analysis

KEGG pathway enrichment analysis was performed on the 521 differentially abundant LC–MS features annotated as putative metabolites, identifying 10 significantly enriched pathways (p < 0.05; Fig. 5). These pathways included biosynthesis of cofactors, neuroactive ligand–receptor interaction, cysteine and methionine metabolism, cAMP signaling, synaptic vesicle cycle, purine metabolism, and related metabolic processes.

Fig. 5.

Differentially abundant metabolitemetabolite enrichment analysis bubble. Notes: The x-axis represents the Rich factor, defined as the ratio of differentially abundant metabolites to all annotated metabolites within a given pathway. Dot size indicates the number of differentially abundant metabolites mapped to each pathway. Pathways with higher Rich factor values reflect a greater proportion of altered metabolites

Several pathways annotated as HIV infection and Flavivirus infection were also enriched, which likely reflects shared metabolic intermediates and signaling components rather than disease-specific viral processes, a common phenomenon in KEGG-based enrichment analyses of metabolomic data.

Across the enriched pathways, a predominance of downregulated annotated features was observed (Fig. 5), suggesting coordinated alterations in metabolic pathways related to energy metabolism, amino acid turnover, and cellular signaling. Given the exploratory nature of this untargeted metabolomics study, pathway enrichment results were interpreted at a system level to provide biological context for the observed metabolic perturbations, rather than as evidence of specific pathway activation or inhibition.

Discussion

This study is the first to characterize metabolic alterations associated with malignant cerebral edema in patients with large vessel occlusion using untargeted red blood cell metabolomics. As LVO is predominantly caused by large artery atherosclerosis and cardioembolism, the metabolic signatures identified in this cohort mainly reflect these two etiological subtypes, which may enhance the internal consistency of the study population.

To determine the most informative biofluid and sampling time point for early MCE-related metabolic changes, a pilot screening was conducted across multiple biofluids and time points. Although differential metabolites were also observed at later stages (days 3 and 7), day-1 RBC samples exhibited the most pronounced and consistent group-level differences while representing the earliest clinically actionable window for MCE risk stratification. An additional observation from the pilot screening was that red blood cells showed more prominent metabolic alterations than plasma or urine. This may be related to the unique biological characteristics of erythrocytes, which are directly exposed to hypoxia, oxidative stress, and systemic metabolic disturbances after acute ischemic stroke [28, 29]. In contrast, plasma metabolites may be partially buffered by protein binding or dilution effects, while urinary metabolites often reflect delayed or cumulative systemic changes. These features may make RBCs more sensitive to early metabolic alterations associated with subsequent malignant cerebral edema, supporting their selection as the primary biofluid for further analysis.

In the expanded cohort, 521 differentially abundant LC–MS features were identified, involving 10 significantly enriched metabolic pathways. From these, 17 metabolites were selected for focused analysis based on a combination of statistical significance (VIP ≥ 1, p < 0.05), effect size (absolute fold change), confident annotation, and biological relevance to ischemic brain injury, including energy metabolism, oxidative stress, inflammation, and membrane integrity. These metabolites have not been previously reported in association with malignant cerebral edema and may therefore provide preliminary insights into MCE-related metabolic dysregulation at the pathway level. Importantly, this selection was intended to highlight representative metabolic changes rather than to define definitive biomarkers, and further validation using targeted metabolomics and authentic standards is required.

Within the lipid-related and neuroactive ligand–receptor interaction pathways identified by KEGG enrichment analysis, PEA was significantly downregulated in the MCE group. PEA is an endogenous lipid mediator noted for its potent anti-inflammatory and neuroprotective effects [30]. It exerts its effects primarily through activation of the PPARα pathway, suppression of astrocyte and microglial activation, and inhibition of NF-κB expression [31]. Following ischemic injury, PEA levels typically increase, contributing to neuroprotection [32]. In traumatic brain injury models, exogenous PEA administration has been shown to reduce CED and improve neurological outcomes [33]. Moreover, serum PEA levels are also positively correlated with NIHSS scores in AIS patients [34]. Accumulating evidence suggests that PEA metabolism (encompassing its endogenous synthesis and degradation) may be impaired in certain pathological conditions [35]. In such contexts, reduced PEA levels may exacerbate disease progression [36]. Therefore, the observed decline in PEA in MCE may reflect impaired biosynthesis, accelerated consumption, or destabilization during the disease course.

The level of SAM, a key methyl donor involved in one-carbon metabolism, was significantly decreased in the MCE group. SAM is involved in DNA methylation, the regulation of oxidative stress, and glutathione synthesis. Its deficiency has been associated with vascular dysfunction, oxidative damage, and inflammation [37]. The accumulation of its downstream product, S-adenosylhomocysteine, is a recognized risk factor for AIS [38]. These findings suggest that disruptions in methionine metabolism may contribute to the pathogenesis of MCE, with SAM serving as a potential metabolic biomarker.

Spermine, a polyamine, plays an essential role in preserving cell membrane integrity and modulating inflammation and intracellular signaling [39]. Elevated spermine has been linked to poor outcomes in AIS [40], and experimental models suggest that exogenous spermine may worsen ischemic injury [41]. Furthermore, a large-scale prospective trial demonstrated that higher serum spermine concentrations independently correlate with poor clinical prognosis [42]. These findings implicate spermine in AIS progression and suggest that targeting spermine synthesis or function may help mitigate ischemic brain injury.

A significant reduction in spermine levels was observed in red blood cells, suggesting that localized depletion may precede systemic alterations. The overexpression of spermidine/spermine N1-acetyltransferase 1 (SAT1), a crucial enzyme in polyamine catabolism, has been implicated in modulating polyamine-related metabolic responses poststroke [43], further implicating polyamine dysregulation in the development of MCE.

ROC analysis demonstrated modest discriminative performance for the highlighted metabolites. Among them, glycyl-L-leucine showed the highest AUC value (0.739), indicating a modest discriminative performance that may warrant further investigation in larger, targeted studies. Glycyl-L-leucine is a dipeptide composed of glycine and L-leucine. While glycine acts as an inhibitory neurotransmitter, its role in AIS remains controversial. Previous research has indicated that glycine may confer neuroprotection by modulating N-methyl-D-aspartate (NMDA) receptor activity, potentially affecting stroke prognosis [44], whereas other studies have reported that glycine may reduce infarct volume in a dose-dependent manner through the modulation of glycine receptors. As a novel dipeptide biomarker, the relevance of glycyl-L-leucine in MCE warrants further investigation.

Furthermore, decreased levels of glutamine and γ-glutamylcysteine indicate impaired glutathione synthesis and disrupted glutamate–glutamine cycling, potentially exacerbating oxidative stress and excitotoxicity and thereby promoting neuronal injury [45, 46]. Additional alterations, including changes in butyrylcarnitine, downregulation of phosphatidylcholine, and upregulation of phosphatidylglycerol, suggest energy metabolic dysfunction and membrane instability, which may compromise blood–brain barrier integrity [47].

Although admission NIHSS scores in the MCE group appeared lower than those reported in some previous studies [12], this difference is primarily attributable to the timing of neurological assessment. In the present study, all patients were enrolled within 24 h of stroke onset, and admission NIHSS scores therefore reflect an early, onset-based evaluation. In contrast, many prior studies—including those summarized by Wu et al.—reported NIHSS scores obtained within 24 h of hospital admission, which may correspond to a later stage after symptom onset. Given the progressive nature of malignant cerebral edema, NIHSS scores assessed at later time points are likely to be higher. Consistent with this interpretation, 24-hour NIHSS scores in our cohort demonstrated a clear separation between the MCE and non-MCE groups and were comparable to values reported in previous studies of large vessel occlusion–related malignant edema.

These considerations underscore the limitations of relying solely on clinical scales to identify patients at high risk for malignant cerebral edema during the early phase after stroke. MCE often develops dynamically after admission, and early biological signals preceding radiological deterioration remain poorly defined. In this context, metabolomic profiling may provide complementary pathophysiological information that is not fully captured by neurological examination or routine laboratory tests.

In the present study, metabolic alterations were detected in red blood cells collected within 24 h of stroke onset, before the occurrence of radiologically confirmed MCE. This suggests that early metabolic disturbances may precede overt cerebral edema and reflect systemic responses related to energy metabolism, oxidative stress, and membrane integrity. Although several metabolites identified here have been linked to stroke severity or neurological outcomes in previous studies, correlations between individual metabolite levels and clinical outcome measures were not formally assessed in this analysis. Accordingly, the findings should be considered exploratory and hypothesis-generating. Future studies integrating targeted metabolomics with longitudinal clinical outcomes are warranted to clarify their prognostic value and potential role in early risk stratification.

Despite identifying multiple potential metabolic mechanisms underlying MCE, this study has several limitations: First, this study was restricted to patients with large vessel occlusion, and the findings may not be generalizable to other stroke subtypes. Second, given the high dimensionality of untargeted metabolomics data, the risk of false-positive findings due to multiple comparisons cannot be fully excluded. Although stringent filtering criteria were applied, the results should be interpreted with caution and require validation in independent cohorts using targeted metabolomics approaches. Third, metabolomic profiles may be influenced by factors such as diet, medications, and inter-individual variability, which could not be fully controlled in this real-world clinical setting. Fourth, patient diagnosis and classification were based on the 2018 Chinese Guidelines for Acute Ischemic Stroke Management, which are routinely used in China. Although these guidelines are fully aligned with the AHA/ASA guidelines in terms of diagnostic criteria and clinical scales, differences in regional clinical workflows may be unfamiliar to some international readers. Fifth, multivariate analyses (PCA, PLS-DA, and OPLS-DA) were used for exploratory pattern recognition. Permutation testing of the OPLS-DA model yielded a negative Q² value, indicating limited predictive performance. Accordingly, these models should be interpreted as exploratory rather than predictive tools, and future studies with larger sample sizes and external validation are needed to improve model stability and clinical applicability. Finally, the pilot phase involved a very small sample size and lacked precise onset-to-sampling time in hours, limiting conclusions regarding temporal metabolic dynamics. In addition, MCE classification was based on follow-up imaging rather than real-time prediction, and the exact timing of MCE onset relative to sampling could not be determined.

Conclusion

In conclusion, untargeted metabolomics profiling of RBCs collected within 24 h of stroke onset revealed distinct metabolic alterations and disrupted pathways associated with subsequent development of MCE in patients with LVO. These findings provide exploratory insights into early metabolic disturbances potentially involved in MCE pathophysiology. However, given the exploratory design and limited sample size, further validation in larger cohorts using targeted metabolomics, as well as mechanistic studies in experimental models, is required to confirm these observations and clarify their biological significance.

Authors’ contributions

Jianping Yu contributed to the paper conception and design. Medical history preparation and data collection were performed by Yuanfeng Zhou, Xin Yang, Bozhi Zhang, and Jiexin Li. Data analysis was conducted by Yuanfeng Zhou. Supervision was provided by Youlin Wu. The first draft of the manuscript was written by Yuanfeng, the revision of the manuscript was completed by Jianping Yu, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the Natural Science Foundation of Chengdu Medical College (No. CYZ18-20), the Education and Teaching Reform Project of Chengdu Medical College (No. JG202007), and the Graduate Student Innovation Project of Chengdu Medical College (No. YCX2022-01-10).

Data availability

The datasets generated during the current study are available in the Open Archive for Miscellaneous Data (OMIX) at the National Genomics Data Center (NGDC), China National Center for Bioinformation, and are publicly accessible under accession numbers OMIX010867 and OMIX010868 at https://ngdc.cncb.ac.cn/omix/release/OMIX010867 and https://ngdc.cncb.ac.cn/omix/release/OMIX010868.

Declarations

Ethics approval and consent to participate

Patients’ capacity to provide informed consent was assessed by the attending neurologist based on clinical status, Glasgow Coma Scale score, and the ability to understand and communicate. Patients deemed capable provided written informed consent themselves; otherwise, written informed consent was obtained from their legally authorized representatives. The study was approved by the Ethics Committee of the First Affiliated Hospital of Chengdu Medical College (Approval No. LXKY00006) and conducted in accordance with the Declaration of Helsinki.

Consent for publication

The manuscript is original, has not been published, and is not under consideration elsewhere. All authors have read and approved the final version and agree to its submission. All necessary permissions from the affiliated institutions have been obtained. Informed consent was obtained from all individual participants included in the study.

Competing interests

The authors declare no competing interests.