Neuroinflammation-associated thalamocortical functional connectivity and cognition in older adults

Abstract

Neuroinflammation has detrimental neurocognitive effects in patients with systemic and neuropsychiatric diseases. Yet, outside of these conditions, it is unclear to what extent neuroinflammation modulates brain network-level function in older adults in the general population and, what, if any cognitive performance issues arise as a result. To evaluate these questions, we compared cerebrospinal fluid (CSF) inflammatory cytokine levels, resting-state functional magnetic resonance imaging, and cognitive performance data in 96 community-dwelling older adults (age ≥ 60 yrs). Principal component analysis yielded three components, accounting for > 50% of the variance in CSF cytokine levels; of which, the first principal component (PC1) accounted for 32% of the overall variance. A multivariate pattern analysis regression model, adjusted for age, sex and in-scanner movement, revealed cytokine PC1 values were inversely associated with functional connectivity between the mediodorsal thalamus and multiple cortical regions associated with the frontoparietal, default mode and dorsal attention functional brain networks (p-FWE = 0.037). For the majority of these cortical regions, functional connectivity with the mediodorsal thalamus was not associated with cognitive performance. In a robust multivariable model, narrative memory performances were positively associated with thalamocortical connectivity to a posterior parietal hub region of the frontoparietal and dorsal attention networks [beta = 0.25 (95% CI; 0.01, 0.49); p = 0.040], but this association did not survive correction for the number of evaluated cognitive domains (Bonferroni-adjusted p < 0.01). Study results reveal that neuroinflammatory CSF cytokine levels were inversely associated with functional connectivity between the mediodorsal thalamus and cortical regions largely involved in frontoparietal, dorsal attention and default mode networks, but the functional consequences of this relationship remain to be determined and will require further investigation to determine if reduced corticothalamic functional connectivity may help account, in part, for common illness-related cognitive complaints.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12974-026-03962-5.

Background

Cerebrospinal fluid (CSF) pro-inflammatory cytokine levels are known to be elevated and associated with alterations in both brain activity and cognition in patients with chronic neurologic disorders [1–3], such as Alzheimer's disease (AD) and multiple sclerosis [4, 5]. Additionally, acute illness or injury induces systemic inflammatory responses characterized by increases in these cytokines, which are thought to underlie ‘sickness behavior.’ Sickness behavior describes a set of cognitive, psychological, and behavioral changes that include decreased arousal, physical activity, and appetite, as well as attention and memory deficits [6].

Compared to healthy controls and patients with non-inflammatory neurologic diseases (e.g., normal pressure hydrocephalus or pseudotumor cerebri), patients with systemic neuroinflammatory conditions (e.g., multiple sclerosis) demonstrate increased CSF levels of interleukin (IL)−1 beta (IL-1β), IL-6, tumor necrosis factor-alpha (TNF-α), and IL-8 [7, 8]. Outside of these neuroinflammatory disease states, many of these CSF pro-inflammatory cytokine levels increase with age [9], and elevated peripheral pro-inflammatory cytokine levels among older adults have been associated with cognitive dysfunction [10, 11]. However, it is unclear which brain regions or networks are implicated in the association between elevated CSF pro-inflammatory cytokines and cognitive dysfunction in community-dwelling older adults. This association is further complicated by variation of cytokine levels by brain region [12, 13], and the dual beneficial and harmful effects of some cytokines on neuronal survival and function [14].

Two mechanistically distinct pathways have been proposed by which pro-inflammatory cytokines may alter brain function—a direct neuronal pathway and an indirect neurovascular pathway. Cytokines exert direct effects on neuronal function through multiple converging molecular mechanisms. Their downstream signaling pathways significantly alter monoamine neurotransmitter action by modifying their synthesis, release and reuptake, and these neurochemical perturbations can result in functional disruption. For example, cytokine-driven activation of the indoleamine-2,3-dioxygenase enzyme shunts tryptophan away from serotonin synthesis towards the kynurenine pathway, depleting serotonin precursor availability and generating neuroactive metabolites that further dysregulate dopaminergic and glutamatergic neurotransmission [15]. Systemic inflammation has been shown to reduce brain functional connectivity within corticostriatal reward circuitry in proportion to elevated circulating serum pro-inflammatory levels, which then mediate downstream impairments in motivation and goal-directed behavior in depressed individuals [16]. At the level of the synapse, elevated inflammatory cytokines impair long-term potentiation and reduce excitatory synaptic transmission in hippocampal circuits; effects that are partially reversible with anti-inflammatory cytokine treatment [17]. Cytokines can further suppress hippocampal neurogenesis by promoting microglial activation, including pro-inflammatory cascades (e.g., TNF-α, IL-6) with consequent deficits in memory and mood [18, 19]. These direct cytokine-mediated effects establish that pro-inflammatory cytokines act directly upon neurons and synapses by disrupting neurotransmitter balance, synaptic plasticity and neurogenesis in ways that collectively degrade the integrity of neural circuits.

Pro-inflammatory cytokines can also alter brain functioning via an indirect mechanistic pathway by which the activation of astrocytes influence downstream effects on neurovascular coupling. Astrocytes, which extend perivascular end feet that simultaneously contact both neuronal synapses and capillary endothelium, are uniquely positioned at the interface of neural activity and local hemodynamics [20]. Astrocytes mediate neurovascular coupling through calcium-dependent release of vasoactive arachidonic acid metabolites, as well as potassium release [21, 22], and astrocytic Ca⁽²⁺⁾ signals have been demonstrated in animal models to couple with both positive and negative blood oxygen level-dependent (BOLD) signal response depending upon the brain state [23]. In the context of neuroinflammation, pro-inflammatory cytokines drive reactive astrogliosis, resulting in disrupted calcium signaling and altered vasoactive mediator release with consequent effect on cerebrovascular tone [24]. Evidence implicating indirect pathway inflammatory action include findings that prolonged systemic lipopolysaccharide-induced inflammation suppresses spontaneous and evoked astrocytic Ca⁽²⁺⁾ responses and attenuates sustained functional hyperemia at the capillary level, supporting that inflammation disrupts astrocyte-dependent neurovascular regulation [25].

The relationship between peripheral pro-inflammatory markers and resting-state functional connectivity (rsFC) has been examined in only a small number of functional magnetic resonance imaging (fMRI) studies, though findings have been regionally and methodologically varied, and inconsistently linked to cognitive outcomes. In a study of older adults (mean age 78), Dev et al. [26] found that higher plasma IL-6 and CRP levels were associated with reduced rsFC between the left parietal cortex and other default mode network (DMN) associated cortical regions, and that greater IL-6 concentration was specifically associated with reduced task-related BOLD signal in the left middle frontal gyrus during n-Back working memory performance. This study provides one of the few demonstrations that peripheral inflammatory burden appears to relate to frontoparietal network disruption and to working memory function in an aging adult cohort [26]. In a chronic pain cohort (urological chronic pelvic pain syndrome), Martucci et al. [3] identified a significant association between plasma IL-6 and reduced rsFC between the posterior cingulate cortex (PCC) and left hippocampus, illustrating that cytokine-related connectivity disruption extends to medial temporal regions implicated in memory function, even outside of primary neurological disease.

Extending these prior observational study findings, experimental endotoxemia studies in healthy adults have provided causal evidence that acute systemic inflammation alters rsFC. Intravenous low-dose lipopolysaccharide (LPS) administration, which reliably elevates plasma IL-6, reduced rsFC among amygdala, insula, default mode and frontoparietal networks, while endotoxin-induced increases in IL-6 specifically predicted reduced coupling relative to placebo between the left thalamus and both the left precuneus and right posterior cingulate cortex (i.e., core DMN hubs) [27]. In depressed patients, where peripheral inflammation is reliably elevated, higher serum CRP levels were associated with reduced rsFC between the ventral striatum and ventromedial prefrontal cortex, with this functional connectivity reduction statistically mediating the relationship between CRP and clinically measured anhedonia [16]. Additionally, Felger et al. [16] found that reduced dorsal striatal rsFC similarly mediated the association between CRP levels and motor slowing, with plasma IL-6, IL-1β, and IL-1 receptor antagonist showing parallel inverse relationships with corticostriatal coupling. Among non-demented older adults followed prospectively, Walker et al. [28] found that a composite peripheral inflammatory index, derived from plasma TNF-α receptor-1 and IL-6, predicted lower rsFC within both the DMN and dorsal attention network approximately six years later, with effects most pronounced in APOE ε4 carriers, though direct relationships with cognitive performance measures were not reported.

Across this literature, two limitations are consistently apparent. First, these studies have exclusively used peripheral blood-based markers, which may not reliably capture CNS neuroinflammatory burden. Second, and most relevant for the present study, the links between inflammatory marker-associated connectivity changes and objective cognitive performance have rarely been formally tested, leaving unanswered whether the observed disruptions in DMN, frontoparietal, and thalamocortical circuits carry measurable functional cognitive consequences. This study addresses both gaps by employing CSF cytokine levels as a direct index of CNS cytokine burden and by examining whether any CSF cytokine-associated rsFC differences correspond to performances on standardized neuropsychological assessment measures among cognitively normal, community-dwelling older adults.

To empirically address these questions, we investigated the relationship between CSF cytokine levels and resting-state functional connectivity in 96 community-dwelling older adults, hypothesizing that elevated CSF pro-inflammatory cytokine levels would be reflected in regional alteration of intra- and inter-network functional brain connectivity. Specifically, we hypothesized that elevated CSF cytokine levels would be associated with reduced rsFC between subcortical and cortical regions involved in cognitive control, including frontoparietal and DMN regions. This hypothesis is based, in part, on prior research demonstrating reduced functional connectivity in frontoparietal, default mode and dorsal attention brain networks in association with elevated peripheral pro-inflammatory cytokine levels [28]. Further, we posited that if CSF cytokine-associated functional connectivity patterns carry any notable functional consequences, then we should expect to see secondary cognitive relevance for any significant rsFC associations independent of the cytokine levels used to detect them.

Methods

Participants

Data were obtained from community-dwelling older adults (age ≥ 60 yrs.) who were scheduled for a non-cardiac/non-intracranial surgery, and who enrolled in two prior observational prospective cohort studies: Markers of Alzheimer’s Disease and Cognitive Outcomes after Perioperative Care (MADCO-PC, clinicaltrials.gov NCT01993836) [29]; and Investigating Neuroinflammation Underlying Postoperative Cognitive Dysfunction (INTUIT, clinicaltrials.gov NCT03273335) [30]. This included 96 participants with complete preoperative baseline CSF cytokine data, neuroimaging data devoid of significant movement or signal artifact (42 MADCO-PC participants and 54 INTUIT participants), and preoperative cognitive assessment data. The MADCO-PC and INTUIT studies enrolled older adults who could read and speak English and who lived within one hour driving distance of Duke University Medical Center. Additionally, both studies excluded prisoners, anyone who was pregnant (due to potential MRI safety concerns), patients taking chemotherapy drugs with known cognitive side effects, and anyone taking anticoagulants that would preclude safe lumbar puncture per clinical guidelines [31]. Participants were also excluded if they were prescribed immunomodulatory or anti-inflammatory drugs that could interfere with the relationship between neuroinflammatory and neurocognitive function.

Participant baseline demographic and clinical characteristics are provided in Table 1. The cohort was slightly over 90% Caucasian, split evenly between men and women with a mean age of 68.3 years old (SD = 5.6). Most of the cohort had hypertension, which matches the known prevalence of hypertension among older adults [32, 33]. The median Mini-Mental Status Exam (MMSE) score was 29, and 99% of the cohort had no mild cognitive impairment (MCI) or AD diagnosis, consistent with epidemiologic data showing that most community-dwelling older adults are cognitively normal [34]. 33.3% of participants (n = 32) were APOE ε4 allele carriers and based on previously published CSF AD biomarker cutoff values [35–38], 20% of the participants were amyloid beta positive (A +) and 5% were total tau-positive (T +). The overall white matter hyperintensity burden was low: over 90% of patients had a Fazekas score of 2 or lower, consistent with normal aging and mild cerebrovascular disease [39, 40].

Table 1.

Participant demographics and structural neuroimaging characteristics

Baseline Demographicsa
Age, in years (SD)	68.32 (5.57)
Race (%)
Black or African American	7 (7.3%)
Caucasian/White	88 (91.7%)
Other	1 (1.0%)
Sex (Male) (%)	48 (50.0%)
Height (cm) (Q1, Q3)	170 [163, 178]
Weight (Kg) (SD)	86.19 (17.66)
BMI [Q1, Q3]	29.63 [25.55, 33.00]
DM	21 (21.9%)
HTN	65 (67.7%)
Neurologic Conditionsb	10 (10.4%)
Psychiatric Conditionsc	29 (30.2%)
Years of Education [Q1, Q3]	16 [13, 17]
MMSE Total Score [Q1, Q3]	29 [27, 30]
MMSE Category (%)
< 20	1 (1.0%)
20–24	3 (3.1%)
25–30	92 (95.8%)
Global Cognitive Index (SD)d	0.14 (0.67)
APOE ε4 Allele carriers (%)e	32 (33.3%)
CSF Aβ+f	19 (20.0%)
CSF Tau+f	5 (5.3%)
Structural Neuroimaging Characteristics
Proportional Gray Matter Volume (SD)g	0.407 (0.027)
Proportional White Matter Volume (SD)g	0.334 (0.023)
White Matter Hyperintensity Burden (SD)h	0.835 (1.044)
Fazekas Scale, ordinal severity rating (SD)i	1.4 (0.8)
0	9 (9.7%)
1	48 (51.6%)
2	27 (29.0%)
3	9 (9.7%)

Values listed as either mean (SD) or median [Q1, Q3]

^aDemographics for participants with complete CSF and functional neuroimaging data (N = 96)

^bNeurologic conditions = epilepsy, movement disorder, neurodegenerative disorders

^cPsychiatric conditions = major depressive disorder, generalized anxiety disorder, and post-traumatic stress disorder. None of the patients had diagnosed psychotic spectrum disorders

^dAverage of standardized cognitive domain z-scores (see Table 1)

^eAPOE genotype was determined, as previously described

^fAβ and Tau positivity determined based upon prior thresholds [41]. CSF Aβ1–42 and tau were measured in the MADCO-PC cohort using the INNO-BIA AlzBio3 platform (Innogenetics; Ghent, Belgium) and in the INTUIT cohort using the Roche Elecsys platform (Roche Diagnostics; Basel, Switzerland), since the AlzBio3 platform was no longer commercially available when we obtained the full INTUIT sample set

^gTotal tissue volume divided by intracranial volume to correct for head size differences among participants

^hWhite matter hyperintensity burden = [FLAIR white matter hyperintensity volume/(white matter volume + FLAIR white matter hyperintensity volume)*100]

ⁱThree participants were missing FLAIR neuroimaging sequence data (N = 93)

Cognitive assessment

Participants’ cognitive abilities were assessed using standardized neuropsychological assessment measures of auditory-verbal memory, immediate and delayed recall of unstructured narrative materials [42, 43] and structured word-list stimuli [44, 45], immediate and delayed recall of visual memory for simple line-drawing stimuli [46], auditory-verbal simple and complex attention (i.e., working memory) [47], information processing speed [47], and executive function (i.e., logical sequencing and task switching) [48]. All tests were administered by clinical research staff who were trained and supervised by a licensed clinical neuropsychologist (JB).

Cognitive domain factor extraction was carried out via principal components method on 14 individual cognitive test variables, followed by oblique (varimax) rotation. Factor loadings (weights) of each cognitive test variable were determined using the solution from a factor analysis conducted on 330 study patients with completely observed cognitive test scores from the parent MADCO-PC and INTUIT studies [49]. Factor retention was guided by a > 80% cumulative variance criterion and scree plot inspection (see Suppl. Figure 1). Full factor loadings, cross-loading, and the inter-factor correlation matrix are provided in the Supplementary Materials (see Suppl. Figure 1, Suppl. Table 1). The 5 cognitive domain factors were retained for comparison with neuroimaging data, and we also generated a summary global cognitive performance metric that reflected the average of the 5 cognitive domain scores (see Table 2).

Table 2.

Cognitive assessment measure raw scores, cognitive domain scores and global cognitive performance abilities

Cognitive Domains and Test Variables	Raw Score Range	Mean (SD)
Unstructured (Narrative) Memory Domain		0.11 (1.06)
RMT Immediate Recall – Gist Score	2–10	7.5 (1.6)
RMT Immediate Recall – Verbatim Score	0–18	11.6 (3.1)
RMT Delayed Recall – Gist Score	0–10	7.1 (1.8)
RMT Delayed Recall – Verbatim Score	0–16	9.9 (3.3)
Structured (Word List) Memory Domain		0.22 (0.84)
HVLT-R Immediate Recall Total Score (Sum of Trials 1–3)	8–35	24.4 (5.7)
HVLT-R Delayed Recall Score	0–12	8.1 (3.4)
HVLT-R Delayed Recognition Discrimination Index Score	4–12	10.6 (1.6)
Visual Memory Domain		0.20 (0.94)
WMS-R Visual Reproduction Immediate Recall Score	0–11	6.7 (2.6)
WMS-R Visual Reproduction Delayed Recall Score	0–11	6.1 (2.7)
Processing Speed/Executive Function Domain		0.08 (0.94)
WAIS-R Digit Symbol Substitution Subtest Score	15–73	44.6 (11.0)
Trails Making Test – Part A Completion Time (sec.)	15–86	32 [26, 41]*
Trails Making Test – Part B Completion Time (sec.)	32–300	87 [67, 114]*
Attention & Concentration Domain		0.09 (1.04)
WAIS-R Digit Span Subtest – Forward Span Score	3–13	7.8 (2.3)
WAIS-R Digit Span Subtest – Backward Span Score	1–14	6.7 (2.3)
Aggregate Cognitive Index Score		0.14 (0.67)

Randt Memory Test (RMT); Hopkins Verbal Learning Test – Revised (HVLT-R); Wechsler Memory Scale – Revised (WMS-R); Wechsler Adult Intelligence Scale – Revised (WAIS-R). Summary statistics presented as mean (SD) or median [Q1, Q3] for individual test raw scores. N = 94 (2 participants with incomplete neurocognition data)

^*Median [1st quartile, 3rd quartile]

Biofluid collection

Lumbar punctures were performed using a 25-g pencil point needle, as previously described [31, 50]. CSF was aliquoted in polypropylene tubes, using low-binding tips and frozen at −80 °C within 1 h of collection. CSF aliquots were maintained at −80 °C until they were thawed in batch for subsequent assay (as described below).

CSF biomarker measurements & analysis

CSF cytokines were quantified by the Biomarker Core Facility at the Duke Molecular Physiology Institute. Samples were randomized across cohorts, and all samples from the same individual were run together on a single plate and assayed in duplicate. A human control CSF sample, pooled from 9 individuals and stored at −80 °C in aliquots was run in duplicate on all plates to assess inter-assay variability. To this end, CSF samples were thawed, pooled, mixed by inversion, and spun at 2000 rpm for 5 min, and aliquots of 190 μL each were made and stored at −80 °C to be run in duplicate on each plate to assess inter-assay variability.

Meso Scale Discovery (MSD) assays (Meso Scale Diagnostics; Rockville, MD) were used to measure the following cytokine levels. MSD VPLEX plates with a twofold CSF dilution were used to measure interleukin-1 beta (IL-1β), interleukin-6 (IL-6), interleukin-7 (IL-7), interleukin-8 (IL-8), interleukin-10 (IL-10), interleukin-16 (IL-16), and tumor necrosis factor beta (TNF-β), and with a fourfold CSF dilution for interferon gamma-induced protein 10 (IP-10), monocyte chemoattractant protein-1 (MCP-1), macrophage inflammatory protein-1 alpha (MIP-1α), and thymus and activation-regulated chemokine (TARC). An MSD UPLEX assay with no dilution of CSF samples was used to measure epithelial-derived neutrophil-activating peptide 78 (ENA-78), granulocyte colony-stimulating factor (G-CSF), growth-regulated oncogene alpha (GRO-α), interferon-inducible T-cell alpha chemoattractant (I-TAC), monocyte chemoattractant protein-2 (MCP-2), macrophage inflammatory protein-3 beta (MIP-3β), stromal cell-derived factor-1 alpha (SDF-1α), and TNF-related apoptosis-inducing ligand (TRAIL). For each analyte, we calculated both the intra-assay and inter-assay coefficient of variation (CV). For analytes detected in ≥ 50% of CSF samples, values below the lower limit of detection (LLOD) were imputed as one-half the LLOD (½ LLOD) prior to principal component analysis, consistent with standard practice for left-censored biomarker data. Analytes with values below the LLOD in more than 50% of samples were excluded from analyses.

Due to the number of analytes measured and their inter-correlations, we performed principal component analysis (PCA) on the aforementioned CSF cytokine data [51]. We calculated summary scores for the first three principal components for each subject by extracting the loading coefficients for the individual cytokines and applying them to z-normed values. PCA was performed in SAS v9.4 (Cary, NC) via proc factor and proc score statements.

Neuroimaging sequences & data acquisition

Participant neuroimaging data were acquired on a 3 Tesla General Electric UltraHigh Performance (UPH) magnet with a 48-channel 60-cm gradient head-coil (100 mT/m constant gradient strength, 250 T/m/s slew rate). We obtained high-resolution, 3D fast-spoiled gradient-echo (FSPGR) T1-weighted imaging (axial oblique, TR 6.996 ms, TE 3.012 ms, TI 400 ms, 256² matrix, 11° flip angle, SENSE factor 2, 1 mm³ isotropic voxels); a T2-weighted fluid-attenuated inversion recovery sequence (FLAIR; axial oblique, TR 11000 ms, TE 147.696 ms, TI 2250 ms, 256² matrix, 111° flip angle, SENSE factor 1, 1 × 1x2 mm voxels); and an eyes-open, visual crosshair fixation resting-state sequence (axial oblique, TR 2500 ms, TE 30 ms, 64 × 128 matrix, 60° flip angle, 3 mm³ isotropic voxels). Two runs of the resting-state sequence were acquired during the same session for a total functional resting-state scan data acquisition time of 12 min per participant.

Neuroimaging data preprocessing

Functional and T1-weighted FSPGR anatomical data were preprocessed in MATLAB (R2020a, release 9.8.0.1873465) using a flexible preprocessing pipeline in CONN (release 21.a) [52] and SPM-12 (release 12.7771), including removal of initial scans, realignment with correction of susceptibility distortion interactions, slice timing correction, outlier detection, direct segmentation, MNI-space normalization, and smoothing. The first 6 volumes (15 s total) in each functional run were removed to account for initial magnetic field inhomogeneities. Functional data were realigned using SPM realign & unwarp procedures [53], where all volumes were co-registered to a reference image (first volume of each resting-state run) using a least squares approach and a 6-parameter (rigid body) transformation [54], and resampled using b-spline interpolation to correct for motion and magnetic susceptibility interactions. Temporal misalignment between different slices of the functional data (acquired in interleaved bottom-up order) was corrected following SPM slice-timing correction procedure [55], using sinc temporal interpolation to resample each fMRI data time series to a common mid-acquisition time. Outlier volumes were identified using ART [56] using conservative movement (framewise displacement > 0.5 mm) and signal artifact (global signal change > 3 SD) thresholds [57, 58]. Participants with significant movement or signal artifact issues resulting in valid data less than 1.5 times the interquartile range below the first quartile (≤ −2.70z) were excluded (n = 6). Mean composite movement values were additionally retained for each participant and used as an additional data quality correction covariate in the primary analytical model.

Functional data were denoised using the CONN toolbox denoising pipeline [59], including the regression of potential confounding noise component effects characterized by white matter and CSF timeseries, motion parameters and their first order derivatives, outlier volumes, run effects and their first order derivatives, and linear trends within each of the two resting-state functional runs. We then performed bandpass frequency filtering of the resting-state data timeseries [60] between 0.008 Hz and 0.09 Hz. CompCor [61, 62] noise components within white matter and CSF were estimated by computing the average blood-oxygen level dependent (BOLD) signal as well as the largest principal components orthogonal to the signal average, motion parameters, and outlier volumes within each subject’s eroded segmentation masks. From the number of noise terms included in this denoising strategy, the effective degrees of freedom of the fMRI signal after denoising were estimated to range from 132.8 to 227.1 (mean 171.4) across all subjects [58].

Functional and anatomical data were normalized into standard MNI space, segmented into grey matter, white matter, and CSF tissue classes, and resampled to 2-mm isotropic voxels via a direct normalization procedure using SPM unified segmentation and normalization algorithm [63, 64] with the default IXI-549 tissue probability map template. Functional data were smoothed using spatial convolution with a Gaussian kernel of 8-mm full width half maximum (FWHM).

Total gray and white matter volumes corrected for intracranial volume were derived from the participants’ segmentation data in SPM12; the relative burden of FLAIR white matter hyperintensities (i.e., white matter hyperintensity volume/total white matter volume) for each participant was collected using the MATLAB Lesion Segmentation Toolbox [65]. Clinical visual ordinal rating of white matter hyperintensity burden using the Fazekas Scale [66] was also adjudicated via blind review, with 96% agreement between 2 independent raters (VU & DZ). See Table 1 for volumetric neuroimaging and sample demographic characteristics.

Neuroimaging data analyses

Voxel-level hypotheses were evaluated using multivariate parametric statistics with random effects across subjects and sample covariance estimated across multiple measurements. Functional connectivity multivariate pattern analyses (MVPA) [67] were performed to characterize the principal axes of heterogeneity in whole-brain functional connectivity across subjects. Individual functional connectivity matrices were derived from bivariate correlation coefficients between BOLD timeseries for each pair of voxels, estimated via subject-level singular value decomposition (SVD) of z-score normalized BOLD signals retaining 64 components per subject. 64 components were retained to ensure stable covariance estimation, while remaining computationally tractable. At the group level, a SVD of the stacked functional connectivity matrices (one row per target voxel, one column per subject) yielded six eigenpatterns representing the dominant axes of connectome heterogeneity across participants, along with their associated eigenpattern-score images characterizing each individual's whole-brain functional connectome state among cortical and subcortical gray matter regions. Eigenpatterns and eigenpattern-scores correspond to the left and right singular vectors, respectively, computed separately for each individual seed voxel.

The six eigenpattern-scores were entered as dependent variables in a whole-brain, voxel-wise MVPA regression model with CSF cytokine principal component 1 (PC1) values as the primary predictor and age, sex, and in-scanner movement as covariates. Preprocessing procedures accounted for BOLD signal noise via white matter and CSF nuisance regressors derived from participant tissue segmentations (see Neuroimaging Data Preprocessing). Statistical inference was conducted at the cluster level using Gaussian Random Field theory [68], applying a cluster-forming threshold of p < 0.001 (voxel-level) and a familywise error-corrected threshold of p < 0.05 (cluster-size).

A follow-up seed-to-voxel regression analysis using the same model parameters and statistical thresholds was conducted to identify contributing foci and their directionality for any regions detected in the primary MVPA F-test [69]. Beta-weights, representing the strength and slope of the estimated linear relationship between CSF cytokine PC1 values and resting-state functional connectivity (rsFC), controlling for age, sex, and in-scanner movement, were extracted for any detected significant seed (i.e., mediodorsal thalamus MVPA cluster) to whole-brain voxel-wise cluster relationships. Any resulting seed to cluster-contributing foci beta-weights were subsequently entered as predictor variables in univariable and multivariable regression models examining global and domain-specific cognitive performances as the dependent variables.

Anatomical area labelling of any significant clusters or foci was based on a 7 mm³ search range of the Talairach Daemon Database using Montreal Neurological Space (MNI)-to-Talairach nonlinear transform coordinates [70]. Canonical functional brain network labelling for any significant regional findings were made by directly comparing spatial overlap with the Yeo 7-network functional brain parcellation atlas [71]. Canonical brain networks labels for significant cluster-contributing foci are provided as the percentage of network representation within any significant cluster-contributing foci (see Table 4) and the number of voxels within these foci (see Fig. 2C).

Table 4.

Mediodorsal thalamocortical functional connectivity inversely associated with CSF inflammatory cytokine principal component (PC1) values

MVPA clustera	Cluster maximab	Cluster extent (k)c	Cluster P-FWE
R./L. Mediodorsal Thalamic Nuclei	+ 10, −20, + 10	93	0.037
Cluster-Contributory Focid		Foci Extent (k)	Foci P-FWE	Foci Brain Network Associatione
Foci 1. Maxima (−40, + 8, + 42 MNI) L. Middle Frontal Gyrus (50%) L. Superior Frontal Gyrus (13%)		1083	< 0.0001	Frontoparietal (50%) Default Mode (42%) Salience (7%) Dorsal Attention (1%)
Foci 2. Maxima (−36, −54, + 40 MNI) L. Lateral Occipital Cortex (37%) L. Angular Gyrus (18%) L. Superior Parietal Lobule (11%) L. Supramarginal Gyrus (9%)		938	< 0.0001	Frontoparietal (64%) Default Mode (27%) Dorsal Attention (9%)
Foci 3. Maxima (+ 40, −56, + 44 MNI) R. Lateral Occipital Cortex (35%) R. Angular Gyrus (33%) R. Superior Parietal Lobule (19%) R. Supramarginal Gyrus (8%)		827	0.00001	Frontoparietal (58%) Dorsal Attention (34%) Default Mode (8%)
Foci 4. Maxima (+ 8, −38, + 32 MNI) R/L. Posterior Cingulate Gyrus (63%) R/L. Precuneus (32%)		693	0.00005	Default Mode (67%) Frontoparietal (15%) Salience (12%) Dorsal Attention (6%)
Foci 5. Maxima (+ 30, + 8, + 44 MNI) R. Middle Frontal Gyrus (63%)		264	0.004	Frontoparietal (97%) Dorsal Attention (3%)
Foci 6. Maxima (+ 36, + 34, + 22 MNI) R. Middle Frontal Gyrus (31%) R. Frontal Pole (10%)		247	0.005	Frontoparietal (82%) Salience (18%)
Foci 7. Maxima (+ 24, + 38, + 36 MNI) R. Frontal Pole (70%) R. Superior Frontal Gyrus (14%) R. Middle Frontal Gyrus (10%)		241	0.006	Default Mode (70%) Frontoparietal (30%)

GLM random-effects, voxel-wise multivariate pattern regression analysis (MVPA), controlling for participant age, sex and mean movement. Statistical thresholds: peak voxel p < 0.001, cluster spatial extent (k_E) threshold p-FWE < 0.05

^aMVPA cluster composed of 58 voxels (62%) covering right medial prefrontal thalamus with centroid MNI (+ 8,−18, + 6), 31 voxels (33%) covering left medial prefrontal thalamus with centroid MNI (−6,−18, + 4)

^bMontreal Neurological Space (MNI) ICBM152 nonlinear 6th generation brain atlas coordinates

^ck represents the number of contiguous voxels within each cluster that exceed p < 0.001 voxel-level significance

^dCluster-contributory foci 1–7 are regions that reduce their functional connectivity to the MVPA cluster (medial dorsal thalamic nuclei) in response to increase in CSF inflammatory cytokine levels. Brain regions listed account for > 5% of total voxels within each foci

^ePercentage of the cluster-contributory foci spatial extent that overlap with Yeo 7-network canonical brain parcellation atlas regions [71] (see Fig. 2C.)

Fig. 2.

Thalamoc or tical Resting-State Functional Connectivity Associations with CSF Inflammatory Cytokine Principal Component Values.

Legend. Whole-brain, voxel-wise multivariate pattern analysis (MVPA) identification of brain regions whose resting-state functional connectivity demonstrate significant association with the first principal component of CSF cytokine values (n = 96). A Yellow/orange regions reflect a significant inverse association with cytokine levels in the mediodorsal thalamic nuclei, adjusted for age, sex and movement [F(6,86); voxel-threshold p < 0.001, cluster-threshold p-FWE < 0.05]. B Brain regions contributing to the inverse association between first principal component of CSF cytokine levels and mediodorsal thalamic resting-state functional connectivity (see Table 4; white outlined regions exceed aforementioned statistical thresholds). C Contributory foci/cluster MNI coordinates and their canonical functional brain network associations, based upon spatial overlap with the Yeo 7-network intrinsic connectivity parcellation atlas (polar scale = number of contributory foci/cluster voxels overlapping with network atlas regions)

Cognitive assessment analyses

First, we assessed for direct associations between cognitive performance, neuroinflammatory burden and rsFC in univariable analyses. Second, amongst any significant univariable relationships (p < 0.05), we utilized multivariable linear regression models with adjustment terms for age, sex, race, APOE4 carrier status, education (yrs.), and diabetes and/or hypertension. Diabetes and hypertension were included in the model, since they have been associated with both neuroinflammation and cognitive dysfunction, and could act as confounders [32, 72]. PC1 values were also included as a covariate in these models to test whether any cytokine-associated MVPA rsFC patterns explain variance in cognitive performances beyond that accounted for by CSF cytokine levels alone, revealing potential general cognitive or domain-specific, cognitively-relevant rsFC components in the overall MVPA pattern.

Results

CSF cytokines

To provide a maximally stable PCA solution, we studied relationships among CSF cytokine levels from all 292 subjects who had these data from the parent MADCO-PC (N = 123) and INTUIT (N = 169) studies. Of the 19 cytokines measured in our panel, 4 were excluded from subsequent analysis (IL1β, TNFβ, MIP1α, and I-TAC) because their levels were above the lower limit of detection in less than half of the patient CSF samples. CSF.

cytokine level distributions were skewed; hence, we measured pairwise correlation between CSF cytokines via Spearman correlation coefficients (r_s). Out of 105 cytokine pairs, most were positively correlated, and 83 were significant after FDR-correction (p < 0.05; see Fig. 1). Among these significant correlations, we identified strong associations (i.e., r_s ≥ 0.70) between IL-8 (CXCL8) and GROa (CXCL1) [r_s = 0.80 (95% CI; 0.75, 0.84)], and between IP-10 (CXCL10) and MCP-2 (CCL8) [r_s = 0.71 (95% CI; 0.65, 0.76)]. CSF IL-7 levels had the most inverse associations with other CSF cytokines (7/14), but all associations were weak in magnitude (r_s > −0.2), and only one association was significant [IL-7 to IL-8; r_s = 0.12 (95% CI; −0.23, −0.003)]. We then conducted principal component analyses of the 15 measured cytokines that were detectable in > 50% of the CSF samples among the 292 subjects. Principal components 1–3 (PC1, PC2, PC3) explained 53% of the variability in cytokine levels (32% for PC1, 11% for PC2, 10% for PC3). Thus, PC1 accounted nearly one-third (32%) of the overall variability among cytokines (i.e. ~ 3 times as much of the variance in CSF cytokine levels as PC2 or PC3), and all the cytokine loadings on P01 were positive (see Table 3). The highest loadings were for IP-10 and MCP-2, and the lowest was for IL-7, which is consistent with our pairwise correlation results (see Fig. 1). This pattern suggests that this first principal component can be thought of as a proxy for the overall burden of CSF pro-inflammatory cytokines and regulatory chemokines, like IL-10.

Fig. 1.

Cross-correlation of CSF cytokine levels in study participants.

Legend. Spearman correlation coefficient (r_s) values are presented in black font for significant correlations (p < 0.05) and grey font for non-significant correlations (p ≥ 0.05). Shading represents the strength of correlation; blue indicates inverse correlations, and red indicates positive correlations

Table 3.

CSF cytokine panel analytes and their principal component analysis loadings

CSF analytes (pg/mL)	Median [IQR]	Range	PC loading
IP-10 (CXCL10)	587.31 [388.08, 789.22]	150.87—2675.19	0.16
MCP-2 (CCL8)	3.77 [3.05, 4.75]	1.67—18.02	0.16
TARC (CCL17)	2.50 [1.96, 3.64]	0.66—15.80	0.15
IL-8 (CXCL8)	50.25 [39.74, 63.69]	19.32—149.58	0.15
IL-10 (CSIF)	0.08 [0.06, 0.11]	0.01—0.36	0.13
GROa (CXCL1)	26.64 [21.53, 33.57]	11.14—90.11	0.13
SDF1a (CXCL12)	1394.11 [1111.60, 1664.23]	558.27—2807.46	0.13
MCP-1 (CCL2)	382.83 [320.07, 474.64]	170.10—703.12	0.12
ENA78 (CXCL5)	9.16 [7.65, 11.03]	3.41—36.38	0.12
IL-16 (LCF)	17.29 [14.72, 21.75]	7.73—81.88	0.11
MIP-3b (CCL19)	159.24 [130.51, 210.81]	32.87—558.61	0.11
IL-6 (BSF2)	1.27 [0.97, 1.83]	0.57—10.52	0.06
GCSF (CSF-3)	2.85 [2.11, 4.08]	0.76—25.55	0.05
TRAIL (TNFSF10)	2.98 [2.35, 3.56]	1.23—6.39	0.04
IL-7	0.76 [0.60, 0.97]	0.33—2.26	0.01

Cytokine levels (pg/mL) presented as median [Q1, Q3], range (min, max), and principal component analysis PC1 loading coefficients; (N = 96). Cytokines are presented in order of highest to lowest loading value on PC1. Four additional cytokines were detected in < 50% of samples (IL1β, TNFβ, MIP1α, and I-TAC); thus, these four cytokines were not included in the principal components analysis

Comparison of CSF cytokines levels with functional neuroimaging data

Next, we examined the relationship between cytokine principal component values and resting-state functional connectivity via a whole-brain voxel-wise multivariate pattern analysis (MVPA) regression model, controlling for age, sex and in-scanner movement. This approach revealed a significant regional association between CSF cytokine PC1 values and resting-state functional connectivity (rsFC) of the mediodorsal thalamic nuclei (MNI cluster maxima 10x,−20y,10z; cluster extent (k) 93; p-FWE = 0.037; see Fig. 2A.; Table 4). In contrast, we found no significant associations between CSF cytokine PC2 or PC3 values and regional rsFC. The mediodorsal thalamic MVPA cluster was composed of 58 voxels (62%) covering the right thalamus with centroid MNI coordinates (+ 8,−18, + 6), and 31 voxels (33%) covering the left thalamus with centroid MNI coordinates (−6,−18, + 4).

Post-hoc seed-to-voxel analyses with the MVPA-identified mediodorsal thalamic cluster as the seed revealed that the mediodorsal thalamus MVPA finding was being driven by rsFC reduction in cortical regions associated with the frontoparietal, default mode, dorsal attention, and salience functional brain networks in association with increasing CSF cytokines from the first cytokine principal component (see Table 4, Fig. 2B). The largest cortical contributions to the cytokine association were observed through reduced rsFC between the mediodorsal thalamic nuclei and the left middle frontal/superior frontal gyrus region (cluster-contributing foci 1, k = 1083; frontoparietal, default mode and salience network-associated; see Table 4, Fig. 2C), followed by three other large regions (i.e., spatial extent > 600 voxels). The second largest cortical region was observed in the left posterior parietal/angular gyrus region (cluster-contributing foci 2, k = 938; frontoparietal, default mode and dorsal attention network-associated; see Table 4, Fig. 2C), followed by a cluster-contributing foci in the homologous right hemisphere posterior parietal/angular gyrus region (cluster-contributing foci 3, k = 827, frontoparietal, dorsal attention and default mode network-associated; see Table 4, Fig. 2C) and a final large cluster in the posterior cingulate/precuneus parietal region (cluster-contributing foci 4, k = 693; default mode, frontoparietal, salience and dorsal attention network-associated; see Table 4, Fig. 2C). The remaining significant cluster-contributing foci (5–7) were smaller in spatial extent (range: 241–264 voxels), but similar to the larger clusters, these foci were predominantly associated with frontoparietal and default mode functional networks.

Comparison of CSF cytokines and functional neuroimaging data with cognitive performance

First, we sought to determine whether CSF cytokine PC1 values were directly associated with global cognitive performance. We did not find any significant direct associations between CSF cytokine PC1 values and global cognitive performance in either the univariable [beta = −0.12 (95% CI; −0.27, 0.02); p = 0.101] or multivariable models [beta = −0.07 (95% CI; −0.19, 0.04); p = 0.219; see methods for multivariable adjustment terms]. Similarly, there were no significant direct associations between CSF cytokine PC1 values and individual cognitive domain scores (all p > 0.05).

We then assessed for any significant associations between the strength of cytokine-associated mediodorsal thalamic rsFC to cluster-contributing foci (see Table 4; Fig. 2C) and global cognitive performance. Multivariable linear regression analyses, which controlled for CSF cytokine PC1 values, age, sex, race, APOE4 carrier status, education, and diagnosis of diabetes and/or hypertension, failed to reveal significant relationships between global cognitive performance and rsFC between the mediodorsal thalamus and posterior parietal hub region associated with frontoparietal and dorsal attention networks [beta = 0.06 (95% CI; −0.07, 0.19); p = 0.347; see Suppl. Table 3].

Analyses by individual cognitive domains identified positive univariable associations between the mediodorsal thalamus and a right angular/supramarginal gyrus region involved in frontoparietal and dorsal attention networks (see Fig. 2C and Table 4; cluster-contributing foci 3) and unstructured narrative memory abilities [beta = 0.25 (95% CI; 0.03, 0.46); p = 0.024; see Suppl. Table 4], as well as structured list-learning memory abilities [beta = 0.18 (95% CI; 0.01, 0.35); p = 0.043; see Suppl. Table 4]. Of these, only the association with unstructured narrative memory abilities remained after robust multivariable regression [beta = 0.25 (95% CI; 0.01, 0.49); p = 0.040; see Suppl. Table 5, Suppl. Figure 2], but this finding did not exceed an adjusted statistical threshold for the total number of evaluated cognitive factor domains (e.g., 5 domains; Bonferroni-correction threshold p < 0.01).

Conclusions

In this cohort of community-dwelling older adults, higher levels of a set of pro-inflammatory cytokines (i.e., PC1) were associated with lower resting-state functional connectivity (rsFC) of the mediodorsal thalamic nuclei with multiple cortical regions, primarily involved frontoparietal, default mode, dorsal attention and salience functional brain networks. We did not observe any significant direct associations between the first CSF cytokine principal component (representing CSF pro-inflammatory cytokine burden) and cognition. Relationships between cytokine-associated mediodorsal thalamocortical functional connectivity and cognitive domain performances were largely absent, but one relationship between thalamocortical to right posterior parietal region rsFC and narrative memory survived robust multivariable adjustment, but not correction for the total number of evaluated cognitive domains. Taken together, our findings appear to support a model in which neuroinflammation (i.e., increased CSF cytokine burden) leads to decreased functional connectivity within thalamocortical/corticothalamic loop interactions with brain regions implicated in working memory, introspective thought and directed attention, but it remains unclear how substantial of a role this cytokine-associated thalamocortical functional connectivity association plays in commonly reported subjective neurocognitive performance decrements during heightened inflammatory states.

Our results appear consistent with prior literature on the vital role of the mediodorsal thalamus in working memory function given the heavy representation of the frontoparietal network in the seven significant corticothalamic regions detected by our neuroimaging analyses (see Fig. 2C). While the mediodorsal thalamus has consistently been linked to working memory function in humans [73–75], subjective working memory impairments have been inconsistently reported among patients with focal mediodorsal thalamus lesions. Rather, patients with mediodorsal thalamic lesions report poorly defined issues with arousal, motivation, and memory [76]. This could be due to unilateral mediodorsal thalamic lesions imparting milder impairment than bilateral damage, variability in subregion-specific functionality and lesion locality in the mediodorsal thalamus [77], and/or a more focal role of the mediodorsal thalamus in working memory-related processes such as the maintenance and modulation of information (i.e., sustained cortical representations and attention control) [78]. The latter account fits with the proposed role of the mediodorsal thalamus in the regulation of cortical networks through the maintenance and extension of frontocortical representations to enable memory, attention control, cognitive flexibility, and decision-making [76, 79].

Our study data showed a nominal positive multivariable association between auditory-verbal episodic memory performance for narrative materials and the strength of mediodorsal thalamus rsFC signal to posterior parietal regions with shared representation from frontoparietal and dorsal attention brain networks (cluster-contributing foci 3; see Table 4, Fig. 2C; see Supplementary Materials). While these findings are in line with prior literature on mediodorsal thalamic roles in attention and executive functioning [80], they were not robust to correction for the total number of evaluated cognitive domains (Bonferroni-corrected p < 0.01).

In addition, the mediodorsal thalamus has cortical and subcortical connections that are often studied in the context of affective pain [81], depression [82], and suicidality (i.e., significant microgliosis among suicide patients) [83], since the mediodorsal thalamus is robustly connected to the limbic system and prefrontal cortex [84, 85]. The mediodorsal thalamus is indeed referred to as part of the limbic thalamus as well as the cognitive thalamus [86, 87]. In humans, stimulation of the mediodorsal thalamus induces a reported generalized sensation of ‘unpleasantness’ and is suggested to be a mediator of pain responses, such as aversion behaviors [88]. The broad interconnectedness of the mediodorsal thalamus to cortical and subcortical regions and pain networks is also richly vascularized and vulnerable to inflammation and injury [89]. Studies have shown that, for instance, focal cortical injury in murine models can lead to intensified secondary neuroinflammation in the thalamus that may outlast initial cortical inflammatory responses [90]. Indeed, inflammatory pain-induced reductions in rsFC between the mediodorsal thalamus and medial prefrontal cortex were associated with working memory impairments in rats, highlighting the impact of pain and inflammation on cognitive behavioral outcomes via rsFC-specific brain network impairments [73]. These findings are also consistent with proposed models, in which modest degrees of neuroinflammation (e.g., as measured by CSF cytokine levels) nonetheless may subtly influence function in older adults by leading to altered rsFC between the mediodorsal thalamus and frontoparietal brain network regions, but this mediation has yet to be formally evaluated.

While prior studies have focused on the relationship between CSF cytokine levels and functional connectivity measurement within specific disease states, relatively few studies have examined these associations in community-dwelling older adults without specific neurologic conditions, psychiatric diseases or syndromes. Prior research efforts that have examined inflammation and fMRI neuroimaging data have largely relied upon peripheral blood cytokine levels. For example, in healthy adults, intravenous peripheral injection of bacterial lipopolysaccharide leads to increased serum IL-6 cytokine levels, which leads to reduced rsFC of the thalamus to the left precuneus and right posterior cingulate gyrus, both of which are strongly associated default mode network brain regions [27]. Outside of these select experimentally-induced endotoxemia studies, results from observational studies of peripheral circulating pro-inflammatory cytokines and brain functional connectivity have been mixed. In an adult surgical patient cohort (N = 58) evaluated at presurgical baseline or postoperatively at 3-months, no statistically significant associations were found between plasma IL-6 or serum C-reactive protein (CRP) levels and brain functional connectivity [91]. Yet, in a larger sample of healthy adult controls (N = 98), plasma IL-6 levels were found to be significantly associated with functional connectivity in the cingulate and mediodorsal prefrontal cortices [92]. These studies demonstrate that midline DMN-associated brain regions, such as the posterior cingulate, show activity changes secondary to increased peripheral inflammation. Our findings build on these studies by demonstrating that a heightened neuroinflammatory state (i.e., higher CSF cytokine burden) is associated with altered functional connectivity between the mediodorsal thalamus and frontoparietal, DMN and dorsal attention network regions.

Separate from mediodorsal thalamus-specific literature, cytokine increases within the central nervous system are generally thought to be a neuroinflammatory mechanism that underlies sickness behavior brought on by viral or bacterial insult, or via anesthesia and surgery [6]. Similarly, neuroinflammation is associated with patient complaints of “brain fog,” a common persistent symptom following viral infections such as COVID-19. Although brain-fog remains ill-defined, it is often characterized by a decreased ability to apply concerted and sustained attention to tasks [93, 94]. Changes in astrocyte function in response to inflammatory cytokines may, in turn, alter neuronal activity patterns within brain networks that underlie cognition. For example, astrocyte activation has been found to impact memory function by altering arousal and attentional vigilance [95], like the finding of working memory impairment after induced inflammatory pain in rats [73], which may have resulted from altered arousal and attentional control.

In summary, our study results demonstrate a significant inverse association between CSF cytokine levels and corticothalamic functional connectivity between the mediodorsal thalamus and regions largely involved in frontoparietal, default mode and dorsal attention functional brain networks. Additionally, the magnitude of functional connectivity relationship between the mediodorsal thalamus and frontoparietal network regions had a nominally positive association with narrative episodic memory abilities, but did not survive correction for the total number of evaluated cognitive domains (see Suppl. Table 5, Suppl. Figure 2). No other potentially notable cognitive performance findings were observed to be associated with thalamocortical rsFC magnitude to the six other cortical regions detected in our central neuroimaging analyses (see Fig. 2C).

This work has several limitations. Patients from MADCO-PC and INTUIT studies who could not undergo MRI scans, such as those with MRI-incompatible implants or devices, could not be included in this work, which may have caused a selection bias that could have impacted our ability to fully characterize the relationship between resting-state fMRI thalamocortical signals and CSF proinflammatory cytokine levels. Further, broad generalization of our findings is limited by the single center setting for the MADCO-PC and INTUIT studies and under-representation of participants from underserved and marginalized groups (i.e., 93% of the study cohort self-identified as Caucasian/white) and those with lower education levels (i.e., the mean years of education was 16 years; range 12–17).

Resting-state functional connectivity differences, such as those reported here, reflect the aggregate output of neurovascular-coupled hemodynamic responses and cannot directly dissociate the two mechanistic pathways by which cytokines may impact brain function (e.g., direct neuronal, indirect neurovascular). Our observed reductions in thalamocortical rsFC to cortical regions involved in frontoparietal, default and dorsal attention brain networks could reflect: (1) direct neuronal and synaptic effects of CSF cytokines on circuit activity (e.g., altered neurotransmitter availability or impaired synaptic plasticity); (2) astrocyte-mediated changes in neurovascular coupling that modify the fMRI BOLD signal independently of underlying neural activity; or possibly more likely, (3) a combination of both direct and indirect effects. In our results, lower cytokine-related corticothalamic functional connectivity was observed in regions perfused by anterior, middle cerebral and posterior artery cerebrovascular territories, suggesting that any indirect cytokine pathway effects on the BOLD signal appear to be independent of any differential large-scale regional perfusion changes. Further, our central cytokine association with rsFC was initially detected in the mediodorsal thalamus from a whole-brain MVPA and not observed in other cortical or subcortical regions. It would be more biologically parsimonious to hypothesize that any direct and indirect CSF cytokine pathway effects on functional connectivity are strongest at the level of the mediodorsal thalamus and may not be widely distributed. Results from animal studies appear to lend more support for towards the mediodorsal thalamus being more preferentially impacted by high cytokine levels than the associated distributed cortical regions found to be involved in the lower thalamocortical rsFC, possibly due to higher microglial density, cytokine receptor expression and signal transduction in the basal ganglia than the cortex [96–98]. Future studies pairing fMRI BOLD with co-acquisition of proxy neuroinflammatory state radioligands via positron emission tomography (e.g., translocator protein 18 kDa; TSPO) or electroencephalographic measurement will likely be needed to mechanistically dissociate these contributions.

Additionally, the novel CSF cytokine panel used in this investigation, particularly the combination of MCP-2, IP-10, and IL-8 as primary contributors to PC1 values, does not have a directly comparable multiplex CSF cytokine profile in publicly available datasets to date. Independent replication of our mediodorsal corticothalamic rsFC findings and any secondary cognitive correlates with comparable CSF inflammatory biomarker data represents an important and necessary next step in supporting the observed relationship between CSF cytokine levels and corticothalamic functional connectivity.

Finally, while the primary cytokine panel principal component was composed mainly of pro-inflammatory cytokines and chemokines (cytokines that induce leukocyte migration), many of these cytokines have multiple cellular effects. Collectively, the CSF biomarker panel selected for this investigation has pro-inflammatory, anti-inflammatory, and immune homeostatic signaling profiles; functional involvement in both innate and adaptive immunity; and demonstrated relationships with myriad neuroinflammatory states, such as post-operative delirium, meningitis, and demyelinating and neurodegenerative diseases.

In sum, our study results appear to support that high CSF cytokine levels are significantly associated with lower resting-state brain functional connectivity within thalamocortical/corticothalamic loops, particularly with hub regions of the frontoparietal, default mode and dorsal attention networks. The functional consequences of this relationship were not supportive of any notable cytokine-related rsFC cognitive impairment in otherwise healthy older adults. However, we did observe a potential positive association between cytokine-related rsFC between the mediodorsal thalamus and right posterior parietal hub regions of the frontoparietal and dorsal attention networks with auditory-verbal episodic memory abilities in a multivariable model, which may warrant further investigation. Furthermore, our results suggest the mediodorsal thalamus could be a potentially sensitive regional functional neuroimaging target by which one could evaluate therapeutic approaches for treating neuroinflammation in older adults, particularly if future efforts combine fMRI and PET data acquisition, which may help address lingering questions about the relative role of direct and indirect cytokine action on functional brain networks and the underlying BOLD signal.

Abbreviations

AD

Alzheimer’s disease

APOE4

Apolipoprotein epsilon 4

ART

Artifact detection tool

BBB

Blood–brain barrier

CCL19

Chemokine (C–C motif) ligand 19

CI

Confidence interval

CRP

C-reactive protein

CSF

Cerebrospinal fluid

CXCL1

C-X-C Motif Chemokine Ligand 1

CV

Coefficient of variation

DMN

Default mode network

ENA-78

Epithelial-derived neutrophil-activating peptide 78

FLAIR

Fluid-attenuated inversion recovery

fMRI

Functional magnetic resonance imaging

FSPGR

Fast-spoiled gradient-echo

FWE

Family-wise error

FWHM

Full width half maximum

G-CSF

Granulocyte colony-stimulating factor

GRO-α

Growth-regulated oncogene alpha

HVLT-R

Hopkins Verbal Learning Test-Revised

INTUIT

Investigating Neuroinflammation Underlying Postoperative Cognitive Dysfunction

IL-1β

Interleukin 1 beta

IL-6

Interleukin 6

IL-7

Interleukin 7

IL-8

Interleukin 8

IL-10

Interleukin 10

IL-16

Interleukin 16

IP-10

Interferon gamma-induced protein 10

I-TAC

Interferon-inducible T-cell alpha chemoattractant

LLOD

Lower limit of detection

MADCO-PC

Markers of Alzheimer’s Disease and Cognitive Outcomes after Perioperative Care

MCI

Mild cognitive impairment

MCP-1

Monocyte chemoattractant protein-1

MCP-2

Monocyte chemoattractant protein-2

MIP-1α

Macrophage inflammatory protein-1 alpha

MIP-3β

Macrophage inflammatory protein-3 beta

MMSE

Mini-Mental Status Examination

MNI

Montreal Neurologic Institute

MRI

Magnetic resonance imaging

MSD

Meso scale discovery

MVPA

Multivariate pattern analysis

PCA

Principal component analysis

PC1

Principal component 1

RMT

Randt Memory Test

rsFC

Resting-state functional connectivity

SDF-1α

Stromal cell-derived factor-1 alpha

SENSE

Sensitivity encoding

SPM

Statistical parametric mapping

TARC

Thymus and activation-regulated chemokine

TE

Echo time

TNF-α

Tumor necrosis factor-alpha

TNF-β

Tumor necrosis factor-beta

TR

Repetition time

TRAIL

TNF-related apoptosis-inducing ligand

WAIS-R

Wechsler Adult Intelligence Scale-Revised

WMS-R

Wechsler Memory Scale-Revised

Authors’ contributions

JB: Study conception and design, data processing, data interpretation, manuscript drafting and final approval TR: Data processing, data interpretation, manuscript drafting MCW: Data processing, statistical analysis, data interpretation, manuscript drafting MRS: Data interpretation, manuscript drafting MD: Data interpretation, manuscript drafting PB: Manuscript drafting VRU: Data processing, manuscript drafting DZ: Data processing KM: Manuscript drafting LMS: CSF AD biomarker assays, data processing, data interpretation, manuscript drafting TW: CSF AD biomarker assays, data processing, data interpretation, manuscript drafting CH: Manuscript drafting CTY: Manuscript drafting HZ: CSF AD biomarker assays, data processing, data interpretation, manuscript drafting KB: CSF AD biomarker assays, data processing, data interpretation, manuscript drafting JH: Cytokine assays, Data acquisition and storage, data interpretation, manuscript drafting HJC: Manuscript drafting HW: Manuscript drafting JPW: Manuscript drafting MB: Acquisition of funding, study conception and design, data interpretation, manuscript drafting and final approval.

Funding

This work was supported by a mentored research grant from the International Anaesthesia Research Society and National Institutes of Health (NIH) grants T32-GM08600, R03-AG050918, and 1K76-AG057022 (to MB). Funding support is also acknowledged from R01-HL130443 (JM, JB), U01-HL088942 (JM, JB), U01-AG050618 (JB), and Duke Anesthesiology departmental funds. Dr Berger also acknowledges additional support from the Alzheimer’s Drug Discovery Foundation, the Duke Claude D Pepper Older American Independence Center (P30AG028716), NIH R01AG073598, and a William L Young neuroscience research award from the Society for Neuroscience in Anesthesiology and Critical Care (SNACC). MJD acknowledges additional funding from NIH K23AG084898 and R01AG073598. MB, MJD, and HW also acknowledge additional funding from NIH P30-AG072958 (to HW). HZ is a Wallenberg Scholar and a Distinguished Professor at the Swedish Research Council supported by grants from the Swedish Research Council (#2023–00356, #2022–01018 and #2019–02397), the European Union’s Horizon Europe research and innovation programme under grant agreement No 101053962, and Swedish State Support for Clinical Research (#ALFGBG-71320).

Data availability

The data and analytical code that support the findings of this study are available from the corresponding author upon reasonable request and completion of the necessary inter-institutional data transfer agreements.

Declarations

Ethics approval and consent to participate

Research activities for both MADCO-PC and INTUIT studies were carried out under the supervision and approval of the Duke Institutional Review Board in accordance with ethical research practices and guidelines, as specified in the Declaration of Helsinki [MADCO-PC (IRB# Pro00045180) approved on 7/12/2013; INTUIT (IRB# Pro00083288) approved on 5/24/2017].

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.