The association between inflammatory markers in blood and cerebrospinal fluid: a systematic review and meta-analysis

Abstract

Background:

Neuroinflammatory processes have been hypothesized to play a role in the pathogenesis of psychiatric and neurological diseases. Studies on this topic often rely on analysis of inflammatory biomarkers in peripheral blood. Unfortunately, the extent to which these peripheral markers reflect inflammatory processes in the central nervous system (CNS) is unclear.

Methods:

We performed a systematic review and found 29 studies examining the association between inflammatory marker levels in blood and cerebrospinal (CSF) samples. We performed a random effects meta-analysis of 21 studies (pooled n=1,679 paired samples) that reported the correlation of inflammatory markers in paired blood-CSF samples.

Results:

A qualitative review revealed moderate to high quality of included studies with the majority of studies reporting no significant correlation of inflammatory markers between paired blood-CSF. Meta-analyses revealed a significant low pooled correlation between peripheral and CSF biomarkers (r=0.21). Meta-analyses of individual cytokines revealed a significant pooled correlation for IL-6 (r=0.26) and TNFα (r=0.3) after excluding outlier studies, but not for other cytokines. Sensitivity analyses showed that correlations were highest among participants with a median age above 50 (r=0.46) and among autoimmune disorder patients (r=0.35).

Conclusion:

This systematic review and meta-analysis revealed poor correlation between peripheral and central inflammatory markers in paired blood-CSF samples, with increased correlations in certain study populations. Based on the current findings, peripheral inflammatory markers are a poor reflection of the neuroinflammatory profile.

1. Introduction

Neuroinflammatory processes have been implicated in the pathophysiology of a large range of brain disorders. Obtaining information about inflammatory processes in the central nervous system (CNS) is challenging, as taking a biopsy of brain tissue in humans is invasive and only occurs when it is strictly needed for the diagnostic work-up of for instance a brain tumor. Cerebrospinal fluid (CSF) is the body fluid in closest proximity to the CNS. Analysis of CSF can provide important information about inflammatory processes within the CNS. Yet the collection of CSF through lumbar puncture (LP) is also invasive and therefore not normally part of the clinical work-up in psychiatric disorders, and only justified for research studies if there is not a less invasive method available. Peripheral inflammatory markers in blood are frequently assessed as alternative biomarkers of CNS inflammation. Unfortunately, it is unclear to what extent peripheral inflammatory marker levels are reflective of the levels in CSF and understanding the association between peripheral and central inflammatory markers is crucial.

There is increasing interest in understanding the role of neuroinflammatory processes in psychiatric disorders. Studies assessing neuroinflammation in the CSF in relation to psychiatric disorders are limited and often have a small sample size. As a result, most studies have assessed inflammatory markers in the periphery in relation to psychiatric disorders. Recent meta-analyses have reported increased peripheral levels of IL-6, TNFα, IL-10, IL-13, and IL-18 in major depressive disorder and of IL-6 and IL-1β in bipolar disorder [6, 7]. Also, plasma IL-6 is elevated in those diagnosed with schizophrenia or at high-risk for psychosis [6, 8, 9], and IL-6, IFNγ, IL-1β, and TNFα are increased in autism spectrum disorder compared with controls [10]. A meta-analysis across psychiatric disorders (schizophrenia, bipolar disorder, and depression) [11] has found patterns of cytokine changes in the CSF similar to those found in meta-analyses of peripheral blood cytokines, however how peripheral inflammation relates to neuroinflammation is unclear. Studies of the association between inflammatory markers in paired blood and CSF have reported inconsistent findings. While upregulated plasma levels of IL-6 were associated with increased IL-6 in the CSF of patients with AD [12], other studies did not find a significant relation between peripheral and central IL-6 levels [13, 14]. For CRP, findings are more consistent, given that this was found to be an indicator of inflammation in both the periphery and the CNS [15].

The extent to which the inflammatory profile of blood is reflective of the CSF profile, and vice versa, is dependent on the complex transmission of inflammatory signals between the brain and the periphery. The CNS is dependent on oxygen and nutrients supplied by blood. The CNS is separated from the vascular system through a specialized epithelium cell lining, forming the blood brain barrier (BBB). The BBB is responsible for maintaining homeostasis of the brain by tightly regulating which molecules and cell types from the periphery can pass into the CNS environment [16, 17]. Entry of peripheral inflammatory cells and proteins, such as cytokines, is restricted by the BBB. Numerous pathologic conditions such as trauma and infection can change BBB permeability [18, 19, 20]. Reduced tightness and increased leakiness of the BBB leads to entrance of cytokines and immune cells into the CNS, which in turn may activate glial cells and induce a central inflammatory response [21, 22]. Besides the BBB, blood and CSF interact at the choroid plexus (CP), a highly vascularized structure in the ventricles where CSF is produced from blood plasma through the blood-CSF (B-CSF) barrier [23]. The B-CSF barrier consists of epithelial cells with tight junctions that allow entry of ions (e.g., Na+, Cl-, HCO3-), glucose and small proteins (including inflammatory markers such as cytokines and chemokines), which causes an osmotic gradient that draws water into the CP, producing the CSF [24, 25, 26]. The CP epithelium is also involved in neuro-immune regulation and recruits inflammatory proteins produced in the CNS to the CSF [27, 28]. Lastly, blood and CSF adjoin at the site of the arachnoid granulations, where CSF is reabsorbed into the blood. This allows inflammatory proteins and biochemical waste products in the CSF to be removed from the CNS [29]. Thus, these carefully regulated mechanisms provide three means for associations between peripheral and central inflammatory markers, namely 1) cytokines and chemokines may leak through the BBB and subsequently into the CSF, 2) cytokines and chemokines may leak through the B-CSF barrier, and 3) cytokines and chemokines from the CSF are reabsorbed into the peripheral blood (Figure 1).

Figure 1. Three mechanisms of interaction between peripheral and central inflammatory markers:

1) the blood brain barrier (BBB), where nutrients are exchanged between the blood and CNS; 2) the blood-CSF (B-CSF) barrier in the choroid plexus where CSF is produced from plasma; and 3) CSF resorption at the arachnoid granulations. Various pathologic conditions can alter BBB and B-CSF permeability. These mechanisms provide three means for associations between peripheral and central inflammatory markers, namely 1) cytokines and chemokines may leak through the BBB and subsequently into the CSF, 2) cytokines and chemokines may leak through the B-CSF barrier, and 3) cytokines and chemokines from the CSF are reabsorbed into the peripheral blood. Figure was created with BioRender.com.

We conducted a systematic literature review and meta-analysis of studies assessing the correlation between inflammatory marker levels in the periphery (plasma and serum) and the CSF. Understanding the relation between peripheral and central biomarkers and characterizing which peripheral markers could be used as a proxy to predict central inflammation will be essential for research and clinical applications.

2. Methods and Materials

2.1. Systematic literature review

This systematic literature review was performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) Statement [30]. PubMed and EMBASE library databases were searched using the terms ((paired OR matched) AND ((cerebrospinal fluid OR CSF OR spinal fluid) and (plasma OR serum OR blood OR peripheral))) AND (inflam* OR cytokine OR interleukin). The search was last updated on September 23rd, 2022. Screening and selection of studies was performed by two independent researchers (F.G. and E.S.). Disagreements were resolved through discussion. A flowchart is shown in Figure 2.

Figure 2.

Flowchart of study selection. 29 papers were included for quality review. 21 studies were included for meta-analyses

2.2. Inclusion and exclusion criteria

Studies were evaluated based on a list of predefined inclusion and exclusion criteria. Studies were included in the structural review if the following inclusion criteria were met: 1) paired investigation of a cytokine or chemokine in plasma/serum and CSF, studying the correlation; 2) original research article published in a peer-reviewed journal; and 3) written in English. Exclusion criteria included: 1) plasma/serum and CSF was collected post mortem; 2) samples were taken as part of a medication study (e.g., clinical trials, pharmacological intervention, with no baseline testing); 3) sample size <10.

2.3. Data extraction

Three independent investigators (F.G., E.S. and B.C.) reviewed full-text articles and extracted the data. The following variables were extracted: number of paired plasma/serum and CSF samples, population characteristics (disease group, age, sex), methods (sample collection time, reason for LP, time between collection and processing, sample processing method, cytokine analysis, peripheral sample type), and results (main outcome, Pearson/Spearman correlation coefficients, p-value, albumin concentration). If correlation coefficients were not reported, corresponding authors were contacted to request the data.

2.4. Structural review

A structural review was performed to determine the quality of 29 included studies. The quality was assessed based on various criteria, including the description of the study population, methods and main findings, as well as the robustness, quality controls and strengths of applied methods and procedures. Studies were assigned a +, +/− or - for each category, according to pre-specified guidelines (Supplementary Table 1).

2.5. Meta-analysis

Twenty-one out of 29 studies were included in the meta-analysis. Eight studies were excluded from the meta-analysis because they did not report a correlation coefficient (n=8). Due to the non-additivity of Pearson and Spearman correlation coefficients, these could not be combined in the same meta-analysis [31]. Performing a Fisher’s z transformation was not recommended due to the violation of the normality assumption of Spearman’s correlation coefficients [32]. The number of studies reporting Pearson correlation coefficient was low, likely due to Spearman being a more commonly used method to assess cytokine correlations due to the generally non-normal distribution of cytokines. Studies reporting Pearson correlation coefficients (N=5) were included in a separate meta-analysis and visualized in Supplementary Figure 1. Further meta-analyses were performed on studies reporting Spearman correlation coefficients (N=16), which allowed for several sensitivity meta-analyses (Figure 3, 4, & 5, Supplementary Table 2&3). Data were analyzed with R software (version 1.3.1093), using the metafor package [33]. Studies were combined using random effects meta-analyses [34]. Random effects models were fitted using the multivariate maximum likelihood (MML), which allows to control for multiple inputs from the same study and has been suggested in the literature as an unbiased and efficient estimator of heterogeneity [35, 36]. Test statistics including 95% binomial confidence intervals were calculated using the metafor package [33]. R code is available upon request. As we hypothesize that transmission of inflammatory proteins between the brain and the periphery and vice versa relies on general mechanisms, which are not specific to particular cytokines or chemokines, we first performed a meta-analysis across all markers. Since various cytokines were assessed in three or more studies, this allowed for sensitivity analyses of individual markers. To assess various sources of heterogeneity, studies were stratified by the following: mean/median age, sex, clinical population, and various technical aspects including the reason for LP, cytokine analysis, peripheral sample type and the time between blood and CSF collection. Cochran’s Q statistic, derived from the chi-square test, was calculated to assess the presence of heterogeneity. High Q-values are indicative of higher variability among studies than would be expected due to randomness. The percentage of variability between studies attributable to heterogeneity was determined by calculating the inconsistency index (I²) [37]. A p-value of α<0.05 or I² > 50% indicates significant heterogeneity and suggests further examination of subgroups [38]. The presence of outlier studies and publication bias were assessed using funnel plots. In line with PRISMA guidelines, cytokines assessed in two studies or less were not analyzed as a subgroup. Data are presented in forest plots as pooled Pearson or Spearman correlation coefficient and the corresponding 95% confidence interval (CI).

Figure 3. Pooled cytokine correlations.

3A. Forest plot of 16 studies (pooled n= 1,338 paired samples) indicating significant low pooled correlation r= 0.21 (95% CI= 0.14:0.27), high heterogeneity (I2=90.3%) and significant variability (Q (df=63) = 842.91, p<0.0001) between studies. Figure 3B. Funnel plot showing outlier studies.

Figure 4. Cytokine correlations of individual cytokines in paired blood and CSF after excluding outlier studies.

Figure 4A. IL-6 correlation in paired blood and CSF. Forest plot of 7 studies (pooled n=507 paired samples) indicating low pooled correlation r= 0.26 (95% CI= 0.19:0.33), low heterogeneity (I2=0%) and low variability (Q (df=8) = 5.614, p=0.6904) between studies. Figure 4B. TNFα correlation in blood and CSF. Forest plot of 4 studies (pooled n=212 paired samples) indicating significant low pooled correlation r= 0.30 (95% CI= 0.17:0.42), low heterogeneity (I2=1.85%) and low variability (Q (df=3) = 2.05, p=0.5618) between studies. Figure 4C. IL-8 correlation in blood and CSF. Forest plot of 3 studies (pooled n=244 paired samples) indicating low pooled correlation r= 0.05 (95% CI= −0.07:0.17), low heterogeneity (I2=13.26%) and low variability (Q (df=3) = 3.6095, p=0.3068) between studies.

Figure 5. Sensitivity analyses.

Figure 5A. Cytokine correlation in blood and CSF in autoimmune disorders. Forest plot of 4 studies (pooled n=154 paired samples) indicating significant moderate pooled correlation r= 0.35 (95% CI= 0.21:0.49), low heterogeneity (I2=0.01%) and small variability (Q (df=3) = 3.14, p=0.369) between studies. Figure 5B. Cytokine correlation in blood and CSF in lumbar puncture for research purposes. Forest plot of 3 studies (pooled n=55 paired samples) indicating significant moderate pooled correlation r= 0.30 (95% CI= 0.06:0.54), no heterogeneity (I2=0%) and small variability (Q (df=2) = 1.005, p=0.6050) between studies.

3. Results

3.1. Literature review

The literature search yielded 1,406 studies (Figure 2). After screening abstracts and full text for inclusion and exclusion criteria, 1,380 studies were excluded due to no paired samples (n=762), not describing a cytokine/chemokine of interest (n=323), not studying correlation (n=128), animal study (n=108), not in English (n=34), post mortem derived plasma/serum or CSF (n=5), no original research (n=15) and having medication or intervention (n=5). Three additional studies were included upon cross-reference check. Twenty-nine studies were included in the qualitative review (Supplementary Table 1) and are summarized in Table 1.

Table 1.

Data extracted from 29 studies included in the qualitative review (a subset of 16 reporting Spearman correlation coefficients were included in the quantitative meta-analysis).

[Reference] Author (year)	N	Cytokine	Correlation Coefficient	p-value	Disease Group	Age	Sex (M, F)	Reason for Lumbar Puncture	Sample Processing Method	Serum or Plasma	Albumin Concentration	CSF/Plasma Outcomes
[65] Agorastos (2019)	N=35	IL-6	NR	NR	Combat veterans with PTSD and without PTSD, HC	Mean (SD) Combat veterans PTSD 27.3 (4.6); without PTSD 31.7 (8.9), HC 30.9 (6.1)	35, 0	Research	ELISA	Plasma	NR	No correlation between plasma and CSF IL-6.
[66] Akilimali (2017)	N=54	IL-4	NR	NR	Cryptococcosis associated immune reconstitution inflammatory syndrome, HC	Median (IQR) Cryptococcosis associated immune reconstitution inflammatory syndrome 32 (28–40.5), HC 34 (28–38)	36, 18 (HC 27, 27)	Clinical	Luminex Assay	Plasma	NR	No correlation of cytokine levels between plasma and CSF.
		IL-10	NR	NR
		IL-1β	NR	NR
		TNFα	NR	NR
		IL-6	NR	NR
		IL-12	NR	NR
		IFNy	NR	NR
		IL-2	NR	NR
		IL-5	NR	NR
		IL-17	NR	NR
		IL-13	NR	NR
		IL-8	NR	NR
		MIP-1β	NR	NR
		G-CSF	NR	NR
		GM-CSF	NR	NR
		IL-7	NR	NR
		IL-18	NR	NR
[63] Baker (2001)	N=19	IL-6	0.29	p>0.05	Veterans (PTSD, HC)	Mean (SD) 41.3 (3.1)	19, 0	Research	ELISA	Plasma	NR	No significant correlation between plasma and CSF IL-6. Negative correlation in the PTSD group and positive in the HC group, and significant difference in correlations between groups.
[67] Basu (2015)	N=64	IL-1β	0.928 (r)	p<0.001	Early-onset neonatal sepsis, HC	Mean gestational age (SD) (weeks) EOS 31.3 (1.5) HC 31.6 (2.0)	33, 31	Clinical	ELISA	Plasma	NR	Significant correlation between plasma and CSF IL-1β and TNFα.
		TNFα	0.935 (r)	p<0.001
[68] Burwick (2019)	N=48	IL-6	0.025	p=0.08	Pregnant people (preeclampsia, hypertension, HC)	Mean (SD) Preeclampsia 30.6 (6.7), hypertension 35.1 (5.0), HC 32.8 (5.5)	48, 0	Clinical	ELISA	Plasma	3.9 (3.0–4.8) in HC 3.9 (3.2–5.8) in chronic or gestational hypertension participants 3.5(2.9–5.1) in preeclampsia patients	No correlation between plasma and CSF IL-6 or TNFα.
		TNFα	−0.14	p=0.35
[69] Casals-Pascual (2008)	N=124	TNF	0.35 (r)	p< 0.001	Cerebral malaria	Median 28.5 (months)	NR	Clinical	ELISA	Plasma	NR	Plasma TNF levels were correlated with TNF concentration in CSF.
[70] Courtioux (2009)	N=42	CXCL-13	0.06	p=0.293	African trypanoso-miasis (HAT) patients (sleeping sickness)	Mean (range) HAT 31 (12–65) Healthy 45 (11–88)	Ratio HAT 4.2 Healthy 1	Clinical	ELISA	Serum	NR	No significant correlation between serum and CSF CXCL-13
[14] Eidson (2017)	N=18	IL-6	NR	p=0.60	PD, HC	Parkinson’s disease 51.8 (7.96), HC 53.3 (11.41)	13, 5	Research	Multiplex immunoassay	Serum	NR	Significant correlation of serum and CSF CRP in both groups, but not IL-6.
		CRP	NR	p=0.02
[71] Ellison (2005)	N=111 (*42)	IL-6*	0.31	p=0.06	Preterm infants with periventricular white matter injury	NR	NR	Clinical	ELISA	Plasma	NR	No significant correlation between CSF and plasma, trends toward a correlation for IL-6 and TNFα.
		IL-8*	0.24	p=0.15
		IL-10*	−0.09	p=0.62
		TNFα*	0.35	p=0.08
[72] Gallo (1991)	N=75	IL-2	NR	p>0.05	MS, non-inflammatory neurological diseases, HC	Range MS (22–57), non-inflammatory neurological diseases NS, HC NS	5, 15	NR	ELISA	Serum	NR	No correlation between serum and CSF IL-2.
[73] Han (2020)	N=115	IL-12	−0.0401 (r)	p=0.8034	Moyamoya disease, HC	Mean (SD) Moyamoya Disease 42.47 (14.53), HC 43.31 (9.66)	62, 53	NR	ELISA	Serum	NR	Significant correlation between plasma and CSF IL-1β and TNFα, but not IL-12.
		IL-1β	0.0499 (r)	p=0.0009
		TNFα	0.422 (r)	p=0.0059
[76] Hirohata (1990)	N=14	IL-6	0.462	p>0.05	Systemic lupus erythematosus with central nervous system involvement	NR	NR	Research	Murine hybridoma cells	Serum	NR	No significant correlation between serum and CSF IL-6.
[74] Hirohata (1997)	N=22	IL-6	0.157	p>0.05	Behçet’s syndrome (with and without central nervous system involvement)	Mean (SD) 50.9 (8.6), 52.4 (14.5)	13, 9	Research	Murine hybridoma cells	Serum	NR	No significant correlation between plasma and CSF IL-6.
[75] Hirohata (2021)	N=101	IL-6	0.3025	p=0.002	Neuropsychiatric systemic lupus erythematosus	Mean (SD) 39.4 (14.6)	12, 89	Clinical	ELISA	Serum	11.32 +/− 1.95 in ACS (acute confusional state) 4.39 +/− 0.70 in non-ACS diffuse NPSLE(neuropsychiatric systemic lupus erythematosus) 3.65 +/− 0.52 in focal NPSLE	Significant correlation between serum and CSF IL-6.
[77] Hopkins (2012)	N=21	IL-6	−0.18 (r)	p>0.05	Subarachnoid hemorrage	Median (Range) 53 (36–72)	11, 14	Research	ELISA	Plasma	NR	Weak correlation between plasma and CSF IL-8, but not IL-6.
		IL-8	0.26 (r)	p>0.05
[78] Laurenzi (1990)	N=19 (*1)	IL-1*	NR	NR	Asymptomatic HIV, HIV with AIDS, MS, Inflammatory neurological disease, Non inflammatory neurological disease, Inflammatory nonneurological disease	NR	NR	NR	ELISA and murine hybridoma cells	Serum	Asymptomatic HIV infected patients: 8, HIV infected patients with AIDS: 2, Multiple Sclerosis patients: 8, Inflammatory neurological disease patients: 9, Non inflammatory neurological disease patients: 0, Inflammatory nonneurological disease patients: 0	No correlation between CSF and serum IL-6 and no differences in patient groups.
		IL-6	NR	NR
[79] Lerche (2022)	N=453 (*261, **347, ***210, ****346, *****245)	IL-6*	M: 0.237	M: p=0.004	PD and HC	Mean (SD) Male HC: 65 (12) Male PD: 66 (9) Female HC: 55 (16) Female PD: 66 (9)	313, 188	Clinical	ELISA	Serum	NR	Significant correlation in CSF and serum in IL-6, IL-4, IL-12p40 and in IL-8 and IL-13 females only.
			F: 0.254	F: p=0.030
		IL-4**	M: 0.400	M: p≤ 0.001
			F: 0.382	F: p≤ 0.001
		IL-12p40***	M: 0.455	M: p≤ 0.001
			F: 0.363	F: p=0.006
		IL-8****	F: 0.288	F: p=0.003
		IL-13*****	F: 0.450	F: p=0.004
[80] Miller (2019)	N=117 (*118)	IFNy	0.07	p=0.465	Pregnant people (MDE and healthy controls)	Mean (SD) MDE 33.1 (5.5), healthy 35.3 (3.7)	0, 117	Clinical	Multiplex bead-based assay panel	Plasma	NR	Significant correlation between plasma and CSF IL-18, IL-23, and TNFα, but not IFNy, IL-6, IL-8 and MCP-1.
		IL-18	0.37	p<0.001
		IL-23	0.31	p<0.001
		IL-6	−0.03	p=.0751
		IL-8	−0.01	p=0.950
		MCP-1*	−0.05	p=0.598
		TNFα	0.21	p=0.026
[14] Okafor (2020)	N=160 (*87, **75)	CD14*	0.078	p=0.47	Cryptococcal meningitis	Median (Range) 35 (28.5–40)	81, 79	NR	Luminex assay	Serum	NR	Significant negative correlation in plasma and CSF CD163, IFNγ and IL-4 at baseline.
		CD163*	−0.249	p=0.02
		G-CSF	−0.003	p=0.97
		GM-CSF	0.033	p=0.68
		IFNγ	−0.214	p=0.007
		IL-10	0.021	p=0.80
		IL-12	0.09	p=0.26
		IL-13	0.056	p=0.48
		IL-17	−0.008	p=0.92
		IL-2	−0.025	p=0.76
		IL-4	−0.232	p=0.003
		IL-5	−0.046	p=0.56
		IL-6	0.021	p=0.79
		IL-7	0.038	p=0.64
		IL-8	−0.034	p=0.67
		IL-1β	−0.042	p=0.60
		MCP-1	0.133	p=0.09
		MIP-1α**	0.01	p=0.93
		MIP-1β	0.117	p=0.14
		TNFα	−0.077	p=0.33
[81] Rota (2006)	N=136	IL-10	NR	NR	AD, vascular dementia, normal pressure hydrocephalus, PD with and without dementia, neurologically HC	Mean (SD) AD 71 (6.03), vascular dementia 68 (5.28), normal pressure hydrocephalus 66 (5.44), PD with dementia 69 (6.24) and without dementia 66 (6.43), neurologically HC 69 (8.60)	81, 55	Clinical	ELISA	Serum	PrAD: 5.52±2.06 PrVD:4.26±1.34 NPH: 4.92±0.91 PD: 5.15±1.86 (d+): 5.03±1.65 (d-): 5.23±2.06 NCONT: 4.70±1.34	No correlation between serum and CSF IL-10 or TGFβ.
		TGFβ	NR	NR
[82] Senel (2010)	N=126	CXCL13	0.372	p=0.067	Neuroborreliosis, systemic borreliosis, Guillain-Barré syndrome, Bell’s palsy, other cranial nerve palsies, cephalgia, bacterial central nervous system infections, and viral central nervous system infections	Median (Range) Neuroborreliosis 58 (32–70), systemic borreliosis 44 (40–56), Guillain-Barré syndrome 40 (27–51), Bell’s palsy 40 (26–67), other cranial nerve palsies 65 (54–66), cephalgia 53 (44–64), bacterial infections 53 (36–69) and viral infections 38 (27–63)	NR	Clinical	ELISA	Serum	Neuroborreliosis: 13.8 (9.7 to 23.3) Systemic borreliosis: 6.6 (4.6 to 8.3) Bacterial infections of the central nervous system: 58.1 (35.5 to 90.2) Viral infections of the central nervous system: 11.8 (8.8 to 16.7) Cephalgia: 4.8 (3.7 to 7.1) Guillain Barré syndrome: 17.4 (6.8 to 23.2) Cranial nerve palsies: 6.2 (5.7 to 8.4) Bell’s palsy: 5.2 (4.4 to 6.0)	No correlation between serum and CSF CXCL13.
[83] Sinha (2015)	N=38	TNFα	0.41	p<0.05	Tuberculous meningitis	Mean (SD) 31.82 (13.19)	26, 12	Clinical	ELISA	Serum	NR	Significant correlation between serum and CSF TNFα
		IL-1β	0.139	p=0.404
[62] Sisay (2017)	N=17	IL-8	0.574	p=0.008	MS	Range 20–60	10, 7	NR	BD Cytometric Bead Array	Serum	NR	Significant correlation between serum and CSF IL-8.
[84] Sjögren (2004)	N=19 (*15)	TGFβ	−0.14	p>0.05	Frontotemporal dementia	Mean (Range) 66.2 (47–80)	6, 13	NR	ELISA	Serum	NR	No correlations between serum and CSF cytokines.
		TNFα*	0.41	p>0.05
[12] Sun (2003)	N=141	IL-6	0.74	p=0.001	AD	Median (Range) 75 (52–85)	45, 96	NR	ELISA	Plasma	6.87(6.44–7.33) 95% CI	Significant correlation between serum and CSF IL-6 and MCP-1.
		MCP-1	0.71	p=0.001
[85] Weller (1991)	N=17	IL-6	0.57 (r)	p<0.03	Neuroborreliosis, HC	NR	22, 12	Clinical	ELISA	Serum	HC: 177 ± 10 Groups are patients with neuroborreliosis Group 1 (first lumbar punctures): 588 ± 45 Subgroup [Group 2A (first control puncture): 668 ± 64, Group 2B (second puncture, after antimicrobial treatment): 412 ± 72]	Significant correlation between serum and CSF IL-6.
[64] Wijeyekoon (2020)	N=22	IFNγ	NR	NR	PD	Mean (SD) 65.4 (7.6)	17, 18	Research	V-plex assay	Plasma	NR	No correlations between plasma and CSF cytokines.
		IL-6	NR	NR
		IL-8	NR	NR
		IL-10	NR	NR
		TNFα	NR	NR
[86] Yuan (2015)	N=85	IL-8	CI: −0.08	CI: p=0.58	HIV infected with impaired cognition and normal cognition	Median (Range) 38 (11–76)	59, 26	Clinical	Milliplex	Plasma	NR	Significant correlation between CSF and plasma G-CSF (NC) and IFNα2 levels.
			NC: 0.19	NC: p=0.29
		G-CSF	CI: 0.17	CI: 0.23
			NC: 0.54	NC: 0.001
		IFNα2	CI: 0.69	CI: p< 0.001
			NC: 0.48	NC: p=0.004
[87] Zin (2010)	N=50	IL-6	NR	p>0.05	Chronic pain	Range 32–85	23, 27	Clinical	ELISA	Plasma	NR	No significant correlation between plasma and CSF IL-6 or IL-10.
		IL-10	NR	p>0.05

AD, Alzheimer’s Disease; CXCL-13, CXC chemokine ligand 13; CD14, cluster of differentiation 14; CD163, cluster of differentiation 163; CI, cognitively impaired; ELISA, enzyme-linked immunoassay; F, female; G-CSF, granulocyte colony stimulating factor; GM-CSF, granulocyte- macrophage colony stimulating factor; HC, healthy controls; HIV, human immunodeficiency virus; IFNα2, interferon alpha-2; IFNγ, interferon gamma; IL-1β, interleukin 1 beta; IL-2 interleukin 2; IL-4, interleukin 4; IL-5, interleukin 5; IL-6, interleukin 6; IL-7, interleukin 7; IL-8, interleukin 8; IL-10, interleukin 10; IL-12, interleukin 12; IL-13 interleukin 13; IL-17, interleukin 17; IL-18, interleukin 18; IL-23, interleukin 23; MIP-1α, macrophage inflammatory protein-1 alpha; MIP-1β, macrophage inflammatory protein-1 beta; MDE, major depressive episode; M, male; MCP-1, monocyte chemoattractant protein-1; MS, Multiple Sclerosis; NC, normal cognition; NR, Not reported; PD, Parkinson’s Disease; PTSD, post-traumatic stress disorder; TGFβ, transforming growth factor beta; TNF, tumor necrosis factor; TNFα, tumor necrosis factor alpha; correlation coefficient, assume Spearman rho unless indicated by (r) = Pearson r; N, number of paired samples per study (asterisk indicates cytokine subgroup N if different from overall study N).

3.2. Qualitative review

Across 29 included studies, 27 inflammatory markers were assessed: IL-6, TNFα, IL-8, IL-1β, MCP-1, IFNγ, IL-4, G-CSF, IL-10, IL-12, IL-13, CXCL13, IL-18, IL-23, CD14, CD163, GM-CSF, IL-17, IL-2, IL-5, IL-7, MIP-1α, MIP-1β, TGFβ, IFNα2, IL-1, IL-12p40 in order of frequency. Of 18 studies that assessed IL-6, the majority reported no significant correlation (n=14), and others reported a strong (n=1) or moderate (n=3) correlation. Similarly, of 11 studies that assessed TNFα, most reported no significant correlation (n=6), four studies reported a moderate or low correlation, and one study reported a strong correlation. IL-8 was assessed in nine studies and most reported no significant correlation (n=7), but two studies reported a moderate or low correlation. IL-1β was assessed in five studies that reported a strong (n=1), low (n=1) or no correlation (n=3). Four studies assessed IFNγ, the majority reporting no correlation (n=3) and one study reporting a low correlation. Three studies of MCP-1 reported a strong correlation (n=1), a low correlation (n=1) and no correlation (n=1). Three studies of IL-4 reported a moderate or low (n=2) and no correlation (n=1). Various studies reported a moderate correlation in IL-18, IL-23, IL-13, IL12p40, and G-CSF. Other cytokines/chemokines showed no significant correlation between plasma/serum and CSF. See Table 1.

Overall, the quality of the 29 studies was rated moderate to high (Supplementary Table 1). However, many studies reported incomplete information on important methodological variables, e.g., the time between paired blood and CSF sample collection, or blood and CSF collection and processing methods. In addition, the quality of applied methods varied, as for some studies the time between paired blood and CSF collection was more than two hours apart, and/or quality control of the cytokine measurements was not performed. Also, statistics were not always reported in full, especially in the case of non-primary outcomes. The 95% confidence intervals of the correlation coefficients were not reported for any of the included studies shown in Table 1.

3.3. Meta-analysis

Twenty-one studies reported a correlation coefficient and were included in meta-analyses. Studies reporting Pearson correlation coefficients (N=5, pooled n=341 paired samples) are visualized in Supplementary Figure 1. Further meta-analyses were performed on studies reporting Spearman correlation coefficients (N=16). The median number of paired samples per study was 45 (range 14–160), with a total of 1,338 paired samples in 16 studies. Studies were stratified into cytokine groups; IL-6 (10 studies, n=925 paired samples), IL-8 (6 studies, n=767 paired samples), TNFα (6 studies, n=420 paired samples), and MCP-1 (3 studies, n=419 paired samples). Studies reported findings in various clinical populations: autoimmune disorders (4 studies, n=154 paired samples), meningitis/CNS infections (3 studies, n=324 paired samples), AD/dementia (2 studies, n=160 paired samples), pregnant people (2 studies, n=165 paired samples), veterans (1 study, n=19 paired samples), HIV (1 study, n=85 paired samples), Parkinson’s disease (PD) (1 study, n= 347 paired samples) and preterm infants (1 study, n= 42 paired samples).

3.3.1. Across cytokines

A random-effects meta-analysis was performed of 16 studies reporting Spearman correlation (pooled n= 1,338 paired samples) to assess the correlation between inflammatory markers in the periphery and the CSF and showed a significant low pooled correlation (r=0.21, 95% CI= 0.14; 0.28, p<0.01). The heterogeneity between studies was high (I²=90.7%) and significant variability was observed (Q (df=62) = 842.01, p<0.001) (Figure 3A). A funnel plot indicated outlier studies (Figure 3B). An additional meta-analysis was performed excluding outliers and indicated significant low pooled correlation (r=0.16, 95% CI= 0.07; 0.24, p<0.05) and significant variability (Q (df=17) = 45.68, p<0.001) between studies (n=7) (Supplementary Table 2).

3.3.2. Individual cytokines

IL-6, TNFα, IL-8, and MCP-1 were measured in three studies or more (Supplementary Table 2). Only IL-6 showed a significant low pooled correlation (r=0.27, 95% CI= 0.13:0.41, p<0.05). Heterogeneity and variability were high for all markers (Supplementary Table 2). Meta-analyses of IL-6, TNFα and IL-8 were repeated after excluding outliers (Figure 4). Both meta-analyses of IL-6 (r=0.26, 95% CI 0.19:0.33, p<0.05) and TNFα (r=0.3, 95% CI=0.17:0.42, p<0.05) indicated significant low to moderate pooled correlation and non-significant variability between studies (Q (df=8) =5.614, p>0.05; Q (df=3) =2.05, p>0.05, respectively), while IL-8 showed low, non-significant correlation (Figure 4; Supplementary Table 2).

3.3.3. Sensitivity analyses

Various sensitivity analyses were performed to investigate sources of heterogeneity. Studies were segregated by median age (median age below 50; median age above 50), clinical population (autoimmune disorders; meningitis/CNS infections), mixed sex population, cytokine detection method (Elisa; bead-based method), same day collection of blood and CSF samples, reason for LP (research; clinical), peripheral sample type (plasma; serum) and publication date after 2010. Results are shown in Supplementary Table 3. Pooled correlation coefficients were highest among participants with a median age above 50 (r=0.46), participants with an autoimmune disorder (r=0.35) and lumbar puncture in a research setting (r=0.3), the latter two showing no heterogeneity and non-significant variability (Q (df=3) = 3.14, p>0.05; Q (df=2) = 1.005, p>0.05), respectively) (Figure 5A&B, Supplementary Table 3). Pooled correlation coefficients were significant and below r=0.3 for other sensitivity analyses, except for participants with meningitis/CNS infections (p>0.05) (Supplementary Table 3). These sensitivity analyses suggest that clinical population, median age, reason for lumbar puncture and cytokine detection method are potential sources of heterogeneity.

4. Discussion

In this systematic review and meta-analysis, we assessed the correlation between peripheral and CNS inflammatory markers, to determine to which extent blood cytokine levels can be used as a proxy for neuroinflammation, since CSF collection is invasive. Quality assessment indicated moderate to high quality of 29 included studies, and pointed out qualitative differences among applied methods including time between blood and CSF sample collection and processing time. The majority of studies reported no significant correlation between blood and CSF cytokine levels, while few reported moderate or low correlations for specific cytokines. A meta-analysis confirmed these findings and showed low pooled correlation of inflammatory markers between blood and CSF.

Although our results suggest that inflammatory markers correlate poorly between the periphery and the CNS, in line with results from previous studies [14, 15], several trends were observed. Pooled correlation estimates were higher in certain clinical groups (patients with autoimmune disorders), age groups (median age above 50), cytokine detection methods (Elisa; plasma), and for specific cytokines (IL-6 and TNFα). These findings could be due to disturbed integrity of the blood-CSF interface in clinical populations and with increased age. Various pathological states are known to cause disturbances to the BBB, and its permeability is affected in e.g., brain injury, multiple sclerosis, Parkinson’s disease and Alzheimer’s disease [23, 24, 39]. Increased BBB permeability can lead to influx of cytokines and chemokines into the CSF, which in turn may activate a central inflammatory response [20, 40]. Consequently, the peripheral inflammatory profile may be more reflective of the CSF in these conditions than in other study populations. Few studies provided measurements related to BBB integrity (e.g. albumin quotient or IgG index), which could provide insight in the integrity of the blood-CSF interface (Table 1).

Interestingly, studies of other signaling molecules show highly correlated levels between CSF and blood. Neurofilament light (NfL) chain protein, for example, is expressed in myelinated axons and has emerged as a biomarker of active CNS neuroaxonal injury. NfL is released into the CSF in response to neuronal injury in various neurological conditions including MS, amyotrophic lateral sclerosis (ALS), AD/PD and brain trauma [41]. A meta-analysis of the correlation between blood and CSF NfL showed a strong correlation (r=0.7) [42]. In addition, impaired BBB integrity was associated with increased blood NfL levels, indicating even higher correlations of NfL between blood and CSF in certain clinical populations [43]. These studies have led to the use of NfL in blood as a promising surrogate marker of CSF NfL and as a biomarker of neuronal damage. Similarly, astroglial markers S100 calcium-binding protein beta subunit (S100B) and glial fibrillary acid protein (GFAP), which play an important role in neuronal survival and the regulation of neuroinflammation, have been proposed as peripheral markers of neuronal damage. Various studies have reported increased release of these proteins into the bloodstream in association with CNS injury or BBB dysfunction [44, 45, 46] and found strong correlations between serum and CSF GFAP (r >0.6) [47]. In addition, research has shown moderate to strong correlations among NfL, S100B and GFAP in the serum and CSF [47, 48, 49]. It should be noted that NfL, S100B and GFAP are highly and primarily produced in the brain [50, 51], although S100B is also expressed outside the CNS in adipocytes [52], whereas cytokines and chemokines are produced in the periphery, as well as by resident brain cells, which may increase in aging and under inflammatory or pathological conditions [22, 53, 54]. Under these circumstances, the periphery may increasingly reflect the inflammatory profile of the CNS, and vice versa, and this may be further influenced by disturbed integrity of the blood-CSF interface (Figure 1) [18, 19, 20]. In addition, the extent to which the blood reflects the CNS may be dependent on the type of inflammatory biomarker [55, 56]. Results of the current meta-analysis indicate that, unlike NfL, S100B and GFAP, which are abundantly present in the CNS, peripheral cytokines cannot be used as a proxy for central inflammation.

In addition to sources of heterogeneity identified in the current meta-analysis, including median age, cytokine detection method, sex, time between blood and CSF sample collection, peripheral sample type and reason for LP, other potential sources are likely further impacting cytokine stability and detection. Methodological differences such as time between paired blood and CSF sample collection, temperature at which samples are stored, freeze-thaw cycles and processing delays affect inflammatory marker levels, and should be reported in publications measuring inflammatory marker levels in frozen samples [57, 58]. Blood cells in peripheral blood samples produce cytokines within two hours after collection, which may affect cytokine measurements [59]. To illustrate, IL-6 levels decrease whilst TNFα levels increase in unprocessed blood samples left at room temperature for a few hours [59]. Also, cytokines degrade more quickly at room temperature compared to temperatures below four degrees [60, 61]. In addition, differences between plasma and serum collection and processing protocols may affect cytokine degradation, e.g., serum is left to clot at room temperature while plasma samples are centrifuged and the supernatant is collected and referred to as plasma. Lastly, generally lower cytokine levels in the CSF compared to plasma pose an additional technical challenge, as these might be more vulnerable to freeze/thaw cycles and other methodological differences, possibly leading to reduced assay performance with lower cytokine levels in the noise range [59, 60]. The importance of consistent sample handling was reflected by the low variability and non-significant heterogeneity among LP’s performed in a research setting compared to a clinical setting. Since details of sample processing and storage and assay performance were not reported in all included studies (Supplementary Table 1), methodological differences may pose additional sources of heterogeneity that could not be studied in this meta-analysis.

Strengths of this review are the inclusion of potential confounders in sensitivity analyses and the pooled sample size. Studies of the association of peripheral and central inflammatory markers in paired blood and CSF samples are limited and sample size was often small and of a specific study population [14, 62, 63]. This study assesses the correlation between inflammatory markers in paired blood and CSF in a large pooled sample across study populations and study designs. Limitations include high heterogeneity among studies, which poses interpretive challenges. In addition, not all potential sources of heterogeneity could be assessed due to incomplete reporting or when assessed in less than three studies. Additionally, it should be noted that the studies included in the current meta-analysis are biased towards neurological cohorts rather than psychiatric conditions. Furthermore, limitations of cytokine detection in the CSF may affect cytokine correlations. In addition, multiple publications from the same group may be based on the same participants, possibly posing the problem of overrepresentation. Lastly, publication bias may cause an overestimation of the pooled correlation coefficients as studies may not have reported non-significant correlations or if they were not within the study aims.

To conclude, studies of inflammatory marker associations between blood and CSF are limited, and pooled correlations are generally low, with increased correlations in certain study populations. Based on our findings, peripheral levels of inflammatory markers are not suitable as a proxy for CNS inflammation.