This systematic review and meta-analysis was conducted in accordance with the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Page et al., 2021). The study protocol was prospectively registered on the International Prospective Register of Systematic Reviews (PROSPERO) under registration number CRD420251240858.

We conducted a comprehensive systematic search across seven databases: PubMed, Embase, the Cochrane Library, Web of Science, China National Knowledge Infrastructure (CNKI), VIP Information (VIP), and Wanfang Data. The search encompassed all literature from database inception until December 3, 2025. The strategy incorporated both controlled vocabulary (e.g., MeSH, Emtree) and free-text terms. Relevant search terms for vascular diseases included “vascular,” “stroke,” “cerebrovascular disorders,” “cerebral infarction,” “brain infarction,” “cerebral hemorrhage,” “cerebral small vessel diseases,” and other related terminology. For cognitive impairment, the search terms included “dementia,” “cognitive impairment,” “cognitive decline,” “cognitive dysfunction,” “vascular cognitive impairment,” and “vascular dementia.” For biomarkers, the search focused on terms like “neurofilament light chain,” “Neurofilament light chain protein,” and “neurofilament proteins,” with the search specifically restricted to studies conducted on plasma, serum, or blood samples. To minimize publication bias and identify all eligible data, we supplemented the electronic search by manually reviewing the reference lists of included studies and relevant reviews, as well as searching conference abstracts and preprint repositories. The detailed search strategy is provided in Supplementary material S1. Two investigators independently performed the initial screening of titles and abstracts against the predefined inclusion and exclusion criteria.

Studies meeting the following criteria were included: (1) studies involving adults (≥18 years) clinically diagnosed with VCI or VaD, or individuals at risk of VCI/VaD; (2) studies assessing the blood (plasma or serum) NfL levels; (3) observational study designs, including cohort studies, case–control studies, or cross-sectional studies; (4) studies providing sufficient data to calculate or extract effect sizes, covering both the VCI/VaD group and the cognitively normal control group. Studies were excluded if they met the following criteria: (1) duplicated publications, reviews, meta-analyses, case reports, or animal studies; (2) studies for which the full text could not be accessed, or key data could not be extracted or calculated; (3) research not available in Chinese or English. Additionally, studies with confirmed or suspected overlapping participant cohorts were excluded to ensure sample independence in the meta-analysis.

Two investigators (FLQ and YQW) independently extracted data from each included study using a standardized, pre-piloted data extraction form. Any discrepancies between the extractors were resolved through discussion and, if necessary, by consulting a third senior researcher. The following information was systematically collected: first author, publication year, study duration, country, study design type, VCI subtype, VCI diagnostic criteria, sample type, NfL detection method, statistical adjustment, age, gender composition, sample size, and NfL levels. To ensure comparability across all studies and to maintain a consistent cross-sectional analytical framework, we extracted blood NfL levels from the first available measurement point (baseline or initial assessment) for all analyses, even if a study reported longitudinal follow-up data. For studies reporting NfL concentrations as median with range or interquartile range, we estimated the corresponding mean and standard deviation using validated statistical methods described by Luo et al. (2018). All included studies reported the key data necessary for meta-analysis in a complete and clear manner. Therefore, no requests for additional data were made to the authors of the included studies.

The methodological quality of each included study was assessed independently by two investigators (FLQ and YQW) using the Newcastle-Ottawa Scale (NOS) (Stang, 2010). This scale assesses studies on three dimensions: selection, comparability, and exposure/outcome assessment, with a total of 8 items and a maximum possible score of 9. In line with common practice, studies scoring 6 or more points were of high quality. Any discrepancies in scoring were resolved through discussion, with adjudication by a third senior researcher if consensus could not be reached.

The statistical analysis for this study was performed using Review Manager 5.4 software. Since blood NfL levels are continuous variables, Weighted Mean Difference (WMD) along with their 95% Confidence Intervals (CIs) were used as the combined effect size. Effect sizes from individual studies were pooled using inverse-variance (IV) weighting. The heterogeneity between studies was assessed using the Chi-square test (Cochran’s Q) and the inconsistency index (I²). If the results indicated low heterogeneity (p > 0.05 or I² ≤ 50%), a fixed-effect model was used for the meta-analysis. Conversely, if high heterogeneity was indicated (p < 0.05 or I² > 50%), a random-effects model was used. The combined effect size was presented using a forest plot, and p < 0.05 was considered statistically significant.

Given the limited number of included studies (n = 13), we conducted exploratory subgroup analyses to explore potential sources of heterogeneity and assess the robustness of the primary findings (Higgins et al., 2022). The pre-specified subgrouping dimensions included study design, VCI subtype, method of NfL quantification, geographical region, control group type, specimen type, and statistical adjustment. Meta-analyses within each subgroup were performed using the inverse-variance (IV) weighted random-effects model. Statistical significance of differences between subgroups was assessed using the Chi-square test, with p < 0.05 considered significant. All subgroup findings should therefore be interpreted with caution as hypothesis-generating.

To evaluate the influence of each included study on the combined effect in the presence of significant heterogeneity, the leave-one-out method was employed in this study.

When the number of included studies was ≥10, Egger’s regression test was performed using Stata 12.0, and a funnel plot was generated using Review Manager 5.4 to assess publication bias (Egger et al., 1997).

We retrieved a total of 1970 articles through systematic searches, including 324 from PubMed, 764 from Embase, 522 from Web of Science, 266 from the Cochrane Library, 17 from CNKI, 61 from Wanfang, and 15 from VIP. After excluding 613 duplicate articles, 150 potentially relevant articles were identified through initial screening of titles and abstracts. Following full-text review and data extraction, 13 eligible studies were ultimately included (Wang Z. et al., 2021; Wang J. H. et al., 2021; Jiang et al., 2022; Trieu et al., 2024; Ming et al., 2022; Sanchez et al., 2024; Chong et al., 2023; Ma et al., 2020; Chua et al., 2023; Lee et al., 2025; Yue et al., 2024; Gaur et al., 2025; Sun et al., 2025), involving 3,716 patients in total. Figure 1 shows the flow diagram of the systematic search and selection process.

Table 1 presents the main characteristics of the studies included. Among them, 9 studies were prospective cohort studies (Wang Z. et al., 2021; Wang J. H. et al., 2021; Jiang et al., 2022; Trieu et al., 2024; Ming et al., 2022; Sanchez et al., 2024; Lee et al., 2025; Yue et al., 2024; Gaur et al., 2025), 3 were cross-sectional studies (Chong et al., 2023; Ma et al., 2020; Sun et al., 2025), and 1 was a case–control study (Chua et al., 2023). The included studies were primarily published between 2020 and 2025, and data were mainly collected between 2014 and 2023. The study participants were primarily elderly, averaging over 60 years of age. Of the 13 included studies, 10 adjusted for potential confounders such as age, sex, and educational level in their analyses, while 3 did not perform such adjustments. The studies mainly came from China (Wang Z. et al., 2021; Wang J. H. et al., 2021; Jiang et al., 2022; Ming et al., 2022; Ma et al., 2020; Yue et al., 2024; Sun et al., 2025) (n = 7), while the remaining studies from the Netherlands (Trieu et al., 2024), Canada (Sanchez et al., 2024; Gaur et al., 2025), Singapore (Chong et al., 2023; Chua et al., 2023), and South Korea (Lee et al., 2025). All studies measured NfL levels in blood, with 10 studies using the Single Molecule Array (Simoa) technology (Wang Z. et al., 2021; Wang J. H. et al., 2021; Jiang et al., 2022; Trieu et al., 2024; Sanchez et al., 2024; Chong et al., 2023; Ma et al., 2020; Chua et al., 2023; Lee et al., 2025; Sun et al., 2025), 2 using Enzyme-Linked Immunosorbent Assay (ELISA) (Ming et al., 2022; Gaur et al., 2025), and 1 using Electrochemiluminescence Immunoassay (ECLIA) (Yue et al., 2024). The detection sample types included plasma (n = 8) and serum (n = 5). The diagnosis of VCI was primarily based on multiple neuropsychological assessments and clinical diagnostic criteria, including but not limited to the Montreal Cognitive Assessment (MoCA), Clinical Dementia Rating (CDR), and vascular dementia-related diagnostic criteria (e.g., National Institute of Neurological Disorders and Stroke—Association Internationale pour la Recherche et l’Enseignement en Neurosciences, NINDS-AIREN; Gorelick criteria; Vascular Behavioral and Cognitive Disorders, VASCOG). The bias risk for the 13 included studies was assessed using the NOS, with a median score of 8 (range: 7–9, Table 2). Among the 9 prospective cohort studies, the NOS scores ranged from 7 to 9 (median 8), the 3 cross-sectional studies had scores from 8 to 9 (median 9), and the 1 case–control study scored 8. According to the widely adopted classification criteria of the scale (total score ≥7 is generally considered high quality), all included studies were regarded as high-quality studies.

A total of 13 studies compared the blood NfL levels between VCI patients (n = 1942) and non-VCI patients (n = 1774). The meta-analysis results showed that the blood NfL levels in the VCI group were significantly higher than in the non-VCI group (WMD = 15.06, 95% CI = [11.41, 18.71], p < 0.00001) (Figure 2), with significant statistical heterogeneity (I² = 93%, p < 0.00001).

The meta-analysis showed high heterogeneity (I² = 93%). To explore the source of heterogeneity, we conducted subgroup analyses (Table 3). First, based on the study design, the subgroup analysis showed that blood NfL levels were significantly higher in the VCI group than in the non-VCI group in prospective cohort studies (WMD = 15.27, 95% CI = [9.98, 20.55], p < 0.00001, I² = 94%), cross-sectional studies (WMD = 10.79, 95% CI = [9.76, 11.82], p < 0.00001, I² = 0%), and case–control studies(WMD = 28.10, 95% CI = [17.21, 38.99], p < 0.00001). However, there were significant between-group differences between the study designs (p = 0.002). Secondly, subgroup analysis based on detection methods and Specimen types showed no statistically significant difference in the effect sizes between studies using Simoa technology (WMD = 15.13, 95% CI = [10.94, 19.31], p < 0.00001, I² = 94%) and those using other detection methods (WMD = 13.98, 95% CI = [3.15, 24.82], p = 0.01, I² = 76%) (p value for subgroup difference = 0.85). Similarly, there was no significant difference in the pooled effect sizes between serum samples (WMD = 14.89, 95% CI = [9.26, 20.52], p < 0.00001, I² = 94%) and plasma samples (WMD = 15.17, 95% CI = [10.03, 20.31], p < 0.00001, I² = 89%) (p value for subgroup difference = 0.94). Additionally, subgroup analyses based on VCI subtypes (post-stroke VCI subgroup (WMD = 19.40, 95% CI = [14.87, 23.93], p < 0.00001, I² = 88%) versus non-stroke VCI subgroup (WMD = 10.28, 95% CI = [6.53, 14.04], p < 0.00001, I² = 81%)), country (China subgroup (WMD = 18.18, 95% CI = [13.43, 22.93], p < 0.00001, I² = 95%) compared to other countries subgroup (WMD = 10.38, 95% CI = [4.91, 15.86], p = 0.0002, I² = 82%)), control group type (vascular risk/disease controls (cognitively normal) (WMD = 21.07, 95% CI = [16.85, 25.28], p < 0.00001, I² = 78%) as opposed to cognitively normal healthy controls (WMD = 9.36, 95% CI = [6.82, 11.90], p < 0.00001, I² = 72%)), and statistical adjustment (adjusted subgroup (WMD = 12.49, 95% CI = [11.67, 13.30], p < 0.00001, I² = 94%) compared to unadjusted subgroup (WMD = 16.85, 95% CI = [14.16, 19.53], p < 0.00001, I² = 70%)) all showed that blood NfL levels in the VCI group were significantly higher than in the non-VCI group, with significant between-group differences (p < 0.05).

To evaluate the stability of the meta-analysis results, we conducted a leave-one-out sensitivity analysis on the blood NfL levels to assess the influence of each study on the pooled WMD. Sensitivity analysis (Figure 3) showed that after removing any individual study, including Ming et al. (2022), Yue et al. (2024), Trieu et al. (2024), Sanchez et al. (2024), Chong et al. (2023), Wang Z. et al. (2021), Ma et al. (2020), Chua et al. (2023), Lee et al. (2025), Gaur et al. (2025), Sun et al. (2025), Wang J. H. et al. (2021), and Jiang et al. (2022). The blood NfL levels in the VCI group were significantly higher compared to those in the non-VCI group, and the pooled effect size and its 95% confidence interval did not show any directional or significant magnitude changes.

We assessed whether there was publication bias in the results of all included studies. The results of Egger’s test (p = 0.8516), Begg’s test (p = 0.7603), and the funnel plot all indicated no potential publication bias (Figure 4).