Neurovascular coupling, functional connectivity, and cerebrovascular endothelial extracellular vesicles as biomarkers of mild cognitive impairment

Abstract INTRODUCTION Mild cognitive impairment (MCI) is a prodromal stage of dementia. Understanding the mechanistic changes from healthy aging to MCI is critical for comprehending disease progression and enabling preventative intervention. METHODS Patients with MCI and age‐matched controls (CN) were administered cognitive tasks during functional near‐infrared spectroscopy (fNIRS) recording, and changes in plasma levels of extracellular vesicles (EVs) were assessed using small‐particle flow cytometry. RESULTS Neurovascular coupling (NVC) and functional connectivity (FC) were decreased in MCI compared to CN, prominently in the left‐dorsolateral prefrontal cortex (LDLPFC). We observed an increased ratio of cerebrovascular endothelial EVs (CEEVs) to total endothelial EVs in patients with MCI compared to CN, correlating with structural MRI small vessel ischemic damage in MCI. LDLPFC NVC, CEEV ratio, and LDLPFC FC had the highest feature importance in the random Forest group classification. DISCUSSION NVC, CEEVs, and FC predict MCI diagnosis, indicating their potential as markers for MCI cerebrovascular pathology. Highlights Neurovascular coupling (NVC) is impaired in mild cognitive impairment (MCI). Functional connectivity (FC) compensation mechanism is lost in MCI. Cerebrovascular endothelial extracellular vesicles (CEEVs) are increased in MCI. CEEV load strongly associates with cerebral small vessel ischemic lesions in MCI. NVC, CEEVs, and FC predict MCI diagnosis over demographic and comorbidity factors.

• Functional connectivity (FC) compensation mechanism is lost in MCI.
• CEEV load strongly associates with cerebral small vessel ischemic lesions in MCI.
• NVC, CEEVs, and FC predict MCI diagnosis over demographic and comorbidity factors.

BACKGROUND
Cognitive dysfunction is a global health concern, with dementia projected to affect 152 million people by 2050. 1 The risk of dementia increases with advanced age, 1 and age-related cellular and molecular alterations underly a driving force in cognitive impairment and dementia development. 2Current treatments for Alzheimer's disease (AD) and related dementias focus on addressing symptoms, with limited options for slowing the progression of early-stage dementia. 3Mild cognitive impairment (MCI) precedes dementia and affects up to 18% of the population worldwide. 4MCI represents the earliest symptomatic stage of pathological cognitive decline, 5 primarily manifesting as mild executive function impairments. 6Nearly one-fifth of older adults' progress from MCI to dementia annually. 7This underscores the need to determine shifts in brain function from healthy aging to MCI to provide mechanistic targets for intervention.
Neurovascular coupling (NVC) is a mechanism that coordinates the communication between parenchymal and vascular cell types to regulate cerebral blood flow (CBF) changes in response to neuronal activity. 8Preclinical models highlight the importance of NVC in sustaining cognitive function. 9Human studies demonstrate a strong association between cognitive performance and NVC responses in healthy older adults. 10Decreased NVC responses have been shown throughout the cortical and subcortical regions in individuals with AD and related dementias. 11Despite the significance of NVC in regulating CBF in response to cognitive demands, its alterations in MCI are not fully elucidated.Functional connectivity (FC) captures the statistical relationships within and between neuronal networks, 12 constituting a critical element in cognition regulation. 13Research has shed light on aberrant changes in FC during rest and its association with neurocognitive conditions. 14However, the strength of relationships and regional alterations in FC during cognitive tasks that differentiate MCI from healthy aging are understudied.
Decreased CBF may precede dementia symptoms and pathologic parenchymal changes (i.e., amyloid-β accumulation, neurofibrillary tangles, atrophy). 15Cognitive impairment in patients with MCI is associated with decreased CBF, supporting hemodynamic alterations prior to the development of AD and related dementias. 16Increased burden of ischemic changes in subcortical white matter is an agerelated change that is exacerbated in patients with MCI. 17 Ischemic lesions are represented by white matter hyperintensities (WMH) on T2 fluid-attenuated inversion recovery (FLAIR) magnetic resonance imaging (MRI), with deep WMHs representing a marker for small vessel disease. 18Higher WMH loads have a relationship with decreased CBF and FC strength in areas of cognitive processing. 19is highlights the negative impact of perfusion deficits in small vessel cerebrovascular accidents on the mechanisms crucial for cognitive function.
2][23][24] It is recognized that the cerebrovascular endothelium plays a critical role in regulating transfer of substances across the bloodbrain barrier to maintain parenchymal homeostasis and preserve cognitive health 25 ; however, the roles played by cerebrovascular endothelial EVs (CEEV) in neurocognitive pathology is yet to be uncovered.
To fully understand the relationship between NVC and FC in individuals with MCI, we employed functional near-infrared spectroscopy (fNIRS) measurements during the n-back working memory paradigm.
Blood was obtained from participants to determine the association of plasma CEEV levels with cerebrovascular small vessel ischemic burden.We hypothesized that impaired NVC responses and decreased FC would classify MCI with high feature importance and that CEEV levels would be elevated in MCI compared to CN. Integration of these physiological concepts will yield significant scientific advancements by (1) uncovering characteristic and functional determinants of MCI, thus bridging the gap for a lack of MCI mechanistic study; (2) improving predictive capability of neurocognitive disorders, leading to proactive clinical care in our rapidly aging population; and (3) providing translationally relevant results that will spur longitudinal clinical trials aimed at preventing cognitive impairment.

Sex as a biological variable
Participants were matched based on age, and upon availability, by sex.Both male and female sexes were included in the study.Sex was included as a biological variable in random Forest to determine the importance of this variable in group classification of MCI.

Blood pressure
Blood pressure was measured in accordance with the 2017 AHA/ACC Hypertension Guideline for Standard Measurement of Blood Pressure.
An average of up to three blood pressure recordings were calculated for each participant and systolic, diastolic, and mean arterial pressure (MAP) were compared between groups (CN vs. MCI).

Neuropsychological testing and working memory paradigm
Participants were administered neuropsychological testing from the NIH Toolbox Cognition Battery as described. 27Tests selected included all fluid cognitive measures (i.e., dimensional change card sort, Flanker Inhibitory Control and Attention, Picture Sequence Memory, List Sorting Working Memory, pattern comparison, and oral symbol digit test) and crystalized ability measure (picture vocabulary test).The cognitive n-back working memory paradigm was administered to participants as described. 27In brief, participants were seated in front of a computer monitor and were instructed to click the left mouse button when the target appeared.For 0-back, participants were instructed to click the left mouse button when they saw a W appear on the screen.Participants were instructed to click the left mouse button when they saw a repeat letter (S-D-A-A) for 1-back, and when every other letter is repeated (X-A-T-A) for 2-back.Duration of each n-back block was 72 s. (number of correct responses) was compared between groups for oral symbol digit test.To determine processing speed for each n-back block, we calculated the average RT of correct responses as described. 27nback performance was measured first by calculating the hit rate ( HIT ) and false-alarm rate ( FA ) as described. 28In this study we used the accu-racy index (d') to compare performance between groups.This score is defined as: d ′ = z Hit − z FA , where z denotes z-transform. 28

fNIRS measurement protocol
Continuous changes in cerebrocortical hemodynamic signals during n-back working memory paradigm were measured by fNIRS as described. 27In brief, participants wore a 128 port head cap (Easycap GmbH, Woerthsee-Etterschlag, Germany) with 16 sources and 16 photodetectors (NIRx Medical Technologies LLC, Glen Head, NY) that covered the frontal cortex according to the international 10-20 system.
Sources emit light in the NIR range, 760 and 850 nm, to detect relative changes in deoxy-(HbR) and oxy-(HbO) hemoglobin.Forty-eight source-detector pairs were separated by 3 cm to sufficiently penetrate the cerebral cortex (∼1.5 cm), and all fNIRS signals were recorded with a sampling frequency of 3.9 Hz.The fNIRS montage is shown in Supplementary Figure 1.

Preprocessing and analysis of fNIRS recordings
fNIRS preprocessing and analysis were implemented in Matlab 2023b (Mathworks, Natick, MA) using custom scripts written by the authors and executed in a pipeline employing the Brain AnalyzIR toolbox (commit 46c645d). 29Participants were excluded if more than 80% of channels covering the prefrontal cortex had poor contact between optodes and the scalp 30 and if any LDLPFC channel had poor contact.Motion artifacts and physiological signal components that do not arise from brain activity were removed as described. 27The modified Beer-Lambert Law was applied to preprocessed optical densities and converted to chromophore (HbO and HbR) concentrations. 31r NVC responses, chromophores were prewhitened to reduce serially correlated effects from physiological and motion artifacts, and slow drifts were removed by a discrete cosine transform-based high-pass filter (0.009 Hz).General Linear Model (GLM) analysis of hemodynamic responses assessed NVC at the subject level, generating subject-specific regression coefficients (β) for each n-back block, chromophore, and brain region.
For FC, prior to application of the modified Beer-Lambert Law, a fifth-order Butterworth filter (0.0045-0.4 Hz) was applied to reduce motion and systemic physiological noise contamination, as described previously. 10Following conversion of optical densities to chromophores, correlated HbO-HbR motion-related fluctuations were removed by correlation-based signal improvement.After this final preprocessing step, neural-evoked HbO and related HbR signal were summed (total hemoglobin [HbT]), as described. 32Pearsoncoefficients (r) were quantified for HbT for the entire duration of each n-back block (72 s) for all channel pairs.Spurious correlations were removed by assigning 0 to negative r values, thus reflecting genuine functional brain region connections with significant (p < 0.05), positive r values.Graph theoretical analysis was used to characterize functional brain networks by regarding each channel as a "node" and each connection as an "edge."Binary and weighted parameters were normalized to the maximum strength or number of connections for each node, respectively, yielding a value between 0 and 1. Calculated FC metrics based on graph theoretical analysis include: connection strength for each node ( w D i ), average connection strength for the whole graph (global connection strength -( w D)), three channel (F3-F5, F3-F1, F3-FC3) LDLPFC averaged connection strength ( w D LDP ), binary number of connections for each node (D i ), average number of connections for the whole graph (global node degree -( D)), and three channel (F3-F5, F3-F1, F3-FC3) LDLPFC averaged number of connections (D LDP ).
For further details, we refer the reader to our previous work for calculations of graph theoretical parameters. 10

2.8
Blood draw and processing hours on ice, protected from light.Up to eight samples were labeled at the same time to reduce the effect of labeling time between the first and the last sample.For each stained sample, a nonlabeled sample was prepared by adding 10 μL of plasma to 90 μL of filtered PBS.
The differences in the concentration and ratios of EVs in plasma samples were measured using ApogeeFlow Micro-Plus flow cytometer (ApogeeFlow, UK) (Supplementary Figure 2A).A solution containing silica and polystyrene beads ranging in size from 80 to 1300 nm (ApogeeMix, cat#: 1527, ApogeeFlow, UK) was used daily for the quality control of the instrument (Supplementary Figure 2B).Samples were measured at 1.5 μL/min flowrate for 120 s.Samples were diluted to the 10,000-20,000 events/s concentration (100-to 400-fold dilutions).For each subject, at least two different aliquots of plasma were measured during two separate measurement sessions.
Data were collected and analyzed using Histogram software (v6.0.117,ApogeeFlow, UK).The nonlabeled and fluorescence minus one (FMO) control samples were used to determine channel spillovers and gating strategy.To assess the plasma concentration of vesicles of cerebrovascular origin, the platelet (CD41 + )-and lymphocyte (CD3 + )-negative events were gated out (Supplementary Figure 2C-E).
Subsequently, the concentration (events/μL) of double-positive vesicles (endothelial-and MAL-positive, CD31 + /CD105 + /CD144 + and MAL + ) was determined for each measured aliquot.EVs with positive expression of pan-endothelial markers, negative labeling of CD3 and CD41, and positive expression of MAL-specifically expressed in central nervous system (CNS) endothelial cells and not in peripheral endothelium 23   mitigated by considering NVC responses as significant at q < 0.05 after controlling the FDR by Benjamini-Hochberg-procedure.Therefore, the pipeline was able to determine group-related differences in NVC responses between groups, as well as determine the effect of task difficulty within group.

Statistical analyses
For prediction modeling, we determined an optimal classification model using a systematic approach that was adopted for data preprocessing, feature selection, model training, and validation.Given the relatively small dataset (n = 34) and the desire for a comprehensive evaluation, the process involved leave-one-out cross-validation (LOOCV) to identify the best combination of features and model.In the process of building the classification model, a comprehensive data preprocessing and feature transformation strategy was employed.To standardize the scale of numerical features and facilitate convergence during the modeling process, using the StandardScaler, we standardized features by removing the mean and scaling to unit variance.For the categorical variables, a one-hot encoding scheme was implemented using OneHotEncoder.This encoding is needed for feeding categorical data to many estimators for machine learning models.The algorithm of choice, random Forest, was selected based on its capability to handle mixed feature types and low-dimensional data effectively.It is noteworthy that this data-driven approach was complemented by a predefined constraint on the model's complexity, specifically limiting the maximum features to three (k = 3).After model selection, feature importance was evaluated, and the final model was further tested by different test-train split runs (n = 100) to ensure a comprehensive assessment of its performance.The aggregated metrics on the test set demonstrated consistent and reliable performance.F1 score was calculated as the harmonic mean of recall and precision.

Study participants
Characteristics and baseline physiological data from participants with MCI (n = 20, 71.

Fluid cognitive performance is diminished in participants with MCI while crystalized abilities are maintained in normal age range
Cognitive decline in participants with MCI is confined to subdomains of fluid cognitive functioning (i.e., not dependent on past learned experience), with crystalized abilities significantly unaffected. 35To confirm these cognitive deficits in our cohort, we administered the NIH toolbox cognitive battery (Figure 2A-C In addition to the NIH toolbox cognitive battery, we administered the n-back working memory paradigm (Figure 3A-B).Figure 3A shows the accuracy (d') during each cognitive task and Figure 3B shows

NVC is impaired during the working memory task in participants with MCI
To determine NVC responses elicited by the n-back task, we assessed changes in oxy-hemoglobin concentration (HbO) in the frontal cortex within and between groups (Figure 2C-G).Two participants with MCI and one participant from CN were excluded from analysis because they did not understand one or more of the tasks.Statistical contrast ((2b + 1b) -0-back)) was used to determine the effect of the more cognitively challenging tasks compared to the baseline, low cognitive demand task.This contrast showed a significant increase in NVC in the LDLPFC in the CN group (n = 19) (Figure 2C) and no change in the LDLPFC in the MCI group (n = 18) (Figure 2D).Group comparison of this contrast showed significantly decreased NVC in the LDLPFC in MCI compared to CN (Figure 2E).All statistical contrasts for within   2G).These data indicate that the LDLPFC, a region largely responsible for working memory performance, 36 has decreased NVC in MCI compared to CN at higher cognitive loads and throughout the n-back paradigm.

LDLPFC functional connections are decreased in participants with MCI during the working memory task
To determine the number of functional connections during different levels of difficulty of working memory n-back task, we calculated local network metrics for each fNIRS channel, local node degree (D loc ), and are presented in Supplementary Figure 5 as group averages.Group comparison of D loc (Supplementary Figure 5A-D) show decreased number of functional connections in the LDLPFC in participants with MCI compared to CN during the second 0-back (0b_2) and 1b (p < 0.05).
Subsequently, we calculated normalized global node degree (D), a global network metric representing the average number of connections adjusted relative to the total possible connections.There was no difference in D between groups (Figure 4A).We averaged D for the entire n-back paradigm, defined as task-averaged D, for each group.We determined that there was no difference in task-averaged D between MCI and CN (Figure 4B).
We calculated normalized LDLPFC node degree (D LDP ), representing the average number of functional connections between the LDLPFC and the rest of the cortical channels.There was decreased  n-back conditions.These data show higher connection strengths in CN localized to within and between regions of the LDLPFC compared to participants with MCI.

LDLPFC strength of functional connections is decreased in participants with MCI during the working memory task
We then calculated w D, a global network metric representing the strength of the average number of significant connections adjusted relative to the maximum connection strength.There was no difference between groups in w D (Figure 3 M).w D were then averaged for the whole n-back paradigm, defined as task-averaged w D, for each group.
We determined no significant difference in task-averaged w D in MCI compared to CN (Figure 4N).
Normalized LDLPFC weighted local node degree ( w ( w D LDP ) was calculated, representing the average strength of significant connections (p < 0.05) between the LDLPFC and the rest of the cortical channels.w ( w D LDP ) was decreased in MCI compared to CN irrespective of task condition (95% CI of mean difference = [0.0168to0.1283], p = 0.0111) (Figure 4O).w ( w D LDP ) were then averaged for the entire n-back paradigm, defined as task-averaged w ( w D LDP ), for each group.These data show a trend of decreased task-averaged w ( w D LDP ) in MCI (n = 18, 0.4127 ± 0.1589) compared to CN (n = 19, 0.5095 ± 0.1393, p = 0.0566) (Figure 4P).To determine connection strength from each channel to all frontal cortical nodes, we calculated local network connection strength metrics, ( w D loc ) (Supplementary figure 5E-H).These data show decreased LDLPFC connection strength in MCI compared to CN in 0b_2 and 1b (p < 0.05).

Participants with MCI have an increased load of CEEVs and are associated with high white matter hyperintensity levels
Ratio (%) and plasma concentration (events/μL) of CEEVs were measured in patients' plasma (Figure 5A, B Current evidence indicates that the prevalence of EVs of larger sizes (>1000 nm) may be indicative of increased apoptotic mechanisms. 37 investigate whether there was an indication for an increase of proapoptotic pathways in the MCI group, we performed analysis of size distribution of CEEVs.There were no statistical differences in size distribution in CN versus MCI at less than 180 nm (predicted Myelin and lymphocyte (MAL) labeling was used to identify endothelial cells of cerebral origin. 23We showed no difference in other cell populations positive for MAL following negative gating of endothelial, lymphocyte and platelet EVs (CN: median = 51935 events/μL, IQR = [31420 to 190500], MCI: median = 109900 events/μL, IQR = [11615 to 476000], p = 0.5944) (Supplementary Figure 6C).
CEEV ratio had a significant positive correlation with Fazekas scale grading in patients with MCI (n = 16, Spearman's rho = 0.7050, p = 0.0025) (Figure 5E).These exploratory analyses determined that CEEVs are associated with WMH burden in participants with MCI and that participants with MCI had a higher level of CEEVs than controls.
We investigated the effects of common hypertension medications (e.g., Ca 2+ antagonists, angiotensin-converting enzyme [ACE] inhibitors) and dyslipidemia medications (e.g., statins) and grouped patients by either hypertension medication or no hypertension medication and dyslipidemia medication and no dyslipidemia medication irrespective of MCI and CN status (Supplementary Figure 6D).There was no difference between either grouping, further substantiating CEEVs as a marker of cognition and cerebrovascular disease in MCI.

Diminished LDLPFC NVC responses, CEEVs, and FC predict MCI diagnosis
To determine if the impaired NVC responses, decreased LDLPFC FC, and increased ratio of CEEVs seen in our MCI cohort can predict group classification, we used feature selection process to identify the F I G U R E 6 Neurovascular coupling (NVC), cerebrovascular endothelial extracellular vesicles (CEEVs), and functional connectivity (FC) are the most informative features in mild cognitive impairment (MCI) classification.We determined an optimal classification model using random Forest integrated with leave-one-out cross-validation (LOOCV).Feature importance scores were assigned to variables, highlighting the model's predictive capability.Variables selected for analyses were neurovascular measures (NVC and FC metrics), CEEVs, and data either known to contribute to progression from aging to MCI, and/or data that had a trend toward a differential distribution between groups.The top three features were left dorsolateral prefrontal cortex (LDLPFC) NVC, CEEV ratio, and task-averaged LDLPFC connection strength ( w ) with importance values of 0.4052, 0.3091 and 0.0926, respectively.These features were selected for the final model and following 100 runs with different test-train splits, the chosen model showed a mean predictive accuracy of 68.00%, SD 16.00%.Individual model evaluation, highlighting its validity, achieved a precision of 88.57%, recall of 85.71% and an F1 score of 85.08%.Random Forest integrated with 34 runs of LOOCV assigned importance scores to each feature with LDLPFC NVC having the highest Gini feature importance value (0.4052) followed by CEEV ratio (0.3091) and w D LDP (0.0926).These scores reflect the contribution of each feature to the model's predictive capability (Figure 6, Supplementary Figure 7).This meticulous methodology, rooted in a data-driven approach with predefined constraints, contributes to the establishment of a reliable and interpretable classification model for small datasets.After feature selection, the three variables with the highest Gini score were chosen for the final model ("LDLPFC NVC," "CEEV ratio," and " w D LDP ").

DISCUSSION
This study identified NVC dysfunction and decreased FC in partic- Reduced CBF from aging to MCI to dementia contributes to cognitive impairment pathophysiology 40 and is one of the earliest events in dementia progression. 41However, limited and mixed evidence exists for changes in task-activated hemodynamics between aging and MCI.
Previous human studies showed no difference between MCI and CN in task-evoked frontal cortex NVC responses. 42,43Contrary to these data, others have reported decreased NVC in the frontal and prefrontal cortex. 44,45Interventional trials to improve NVC and cognition are limited, albeit positive results that showed enhanced cognitive function and increased NVC responses after cocoa supplementation in participants with baseline impairment. 46The lack of intervention may be ascribed to the mixed results in human studies, highlighting the Throughout the entire n-back task, NVC dysfunction was localized to the LDLPFC in MCI compared to CN.These findings led us to determine the number, and strength of functional connections from the LDLPFC to the rest of the frontal cortex.The LDLPFC is an essential area for working memory performance and executive function regulation. 36In previous studies assessing FC of the PFC in participants with MCI, both increases 47 and decreases were seen during cognitive tasks. 48Increases were interpreted as a compensatory action and decreases as a failure of the compensation mechanism.We report decreased task-evoked global and LDLPFC number, and strength of functional connections in participants with MCI compared to CN, supporting a failure of compensation for an already impaired neurovasculature.In our recent work assessing NVC and FC changes in aging compared to young adults, we showed that in aging, task-evoked NVC was decreased, and FC was increased, with a significant negative correlation. 10 patients 23 prior to our study.The previous work from Mazzuco et al.
showed increased plasma concentration of CEEVs in MS patients compared to healthy controls. 23Because the endothelium secretes more EVs when damaged or injured, our findings, in line with Mazzuco's previous work, reveal CEEV load as a potential marker of disease in CNS conditions.We also investigated the relationship between CEEVs, functional neuroimaging results, and WMH load.We determined that the ratio of CEEVs from the total EEV population has a significant and strong positive correlation with WMH burden, providing evidence of CEEV load as a potential marker of small vessel ischemic damage.In addition, current evidence indicates that the prevalence of EVs of larger sizes (>1000 nm) may be indicative of increased apoptotic mechanisms. 37EVs are classified into three categories based on size, exosomes (30-150 nm), microvesicles (100-1000 nm), and apoptotic bodies (1000-5000 nm), and these subtypes differ in their biogenesis, release, intravesicular cargo, and function. 53However, criteria based on size overlaps between subtypes and is not exact in classifying EVs. 53th exosomes and microvesicles are involved in cell-to-cell communication and diseased or dysfunctional cells package their cargo into these vesicles, which may have a role in dysfunction of nearby cellular systems. 53While we determined that concentration and ratio of CEEVs were increased in MCI compared to CN, there were no statistical differences in CEEV size ranges between the groups.Current findings in EV literature show exosome and microvesicle molecular cargo is associated with cognitive disease states, 54,55 potentially due to uptake from surrounding cells, compared to large, apoptotic bodies. 56This postulation is congruent with our distribution findings showing a trend of increased exosome and microvesicle distribution in MCI compared to CN.
Preclinical studies have identified that preserving the health of the cerebrovascular endothelium improves cognitive function. 57Here we show molecular and structural MRI correlations in MCI patients that may provide evidence of cerebrovascular endothelial dysfunction.Future directions to elucidate MCI molecular alterations is to assess the cargo (i.e., protein, miRNA) within CEEVs.Recent evidence has revealed that in early-onset MCI patients, exosome protein cargo correlated with cerebral spinal fluid tau levels, WMHs, brain atrophy and cognition scores. 58,59Addressing alterations in molecular cargo of CEEVs will uncover pathways for how increased concentration of CEEVs may cause cerebrovascular damage, and provide insight into new, mechanistically targeted therapy.
F I G U R E 7 Summary of pathologic progression from aging to mild cognitive impairment (MCI).In older age, neurovascular coupling (NVC) is impaired and there is an increase in global functional connectivity (FC) that acts as a memory/cognitive compensation mechanism.We show in MCI, NVC is further decreased, and the FC compensatory response is lost, which classifies MCI with high precision and validity.Cerebrovascular endothelial extracellular vesicles (CEEVs) were elevated in MCI patients compared to controls and significantly correlated with MCI patient small vessel ischemic damage, cognition and FC, and was the second most important feature in MCI classification, potentially implicating CEEVs in the pathogenesis of MCI.
Cross-sectional human physiological studies always pose a challenge in balancing groups with respect to comorbidities.While groups were not matched for comorbidities, no significant differences between groups were seen.Comorbidities that had higher prevalence in MCI compared to CN, such as hypertension and depression, were added to the MCI classification model and showed low feature importance.An additional challenge in human studies is balancing the groups for demographic factors, particularly those that influence brain structure and function.While all participants were not both ageand sex-matched, 15 of the 20 age-matched controls were also sex matched to participants with MCI.We showed no significant difference in education levels in Fisher's exact test when splitting groups between bachelor's degree and higher versus less than a bachelor's degree of educational attainment.However, there were more higher educational achievers in the CN group compared to MCI, but not significantly different.Along with our findings of FC, we suggest the assessment of structural connectivity by diffusion tensor imaging in future studies.
This will aid in unraveling the subcortical white matter neurodegenerative processes that associate with functional alterations in cognitive impairment.In this study, we determined intricate shifts in neurovascular brain function, as well as a potentially novel biomarker that needs further investigation of protein and miRNA cargo for pathway anal- -dwelling older adults and participants with MCI were recruited in this cross-sectional observational study at the University of Oklahoma Health Sciences Center (OUHSC).All screening and data collection were conducted in the Translational Geroscience Laboratory of the Center for Geroscience and Healthy Brain Aging at the OUHSC.Participant screening was performed to determine eligibility prior to participation.Inclusion criteria consisted of age > 50 years old, Clinical Dementia Rating (CDR) equal to 0.5 and/or Montreal Cognitive Assessment (MoCA) <26 and ≥19 for the MCI group.Older adults were included in the CN group if there was no clinical diagnosis of MCI or any cognitive impairment, and MoCA ≥ 26.Participants with chronic conditions (e.g., type 2 diabetes, hypercholesterolemia, hypertension, major depressive disorder) that were pharmacologically uncontrolled, current or prior cerebrovascular complications (e.g., large vessel ischemic stroke with chronic functional impairment), neurodegenerative diseases, and any medical conditions or functional impairments which, in the opinion of the research and clinical neurology team, would render participants unable to complete the study were excluded from the study.All participants except three were right-handed.All participants refrained from consuming caffeine on the day of measurement.Participants were enrolled into the study after signing informed consent.

Fully
corrected T-score, which compares the test taker score to the NIH toolbox nationally representative normative sample and adjusts for demographic variables (i.e., age, sex, race/ethnicity, educational attainment), was used for group comparison.Fully corrected T-scores have a normative, demographic-adjusted population mean of 50 and SD of 10.A composite fully adjusted scale score for fluid cognition comprises dimensional change card sort, Flanker Inhibitory Control and Attention, Picture Sequence Memory, List Sorting Working Memory, and pattern comparison and was used to compare groups overall fluid cognitive ability in multiple cognitive domains.In addition, raw score

Fisher's exact 1 +
test, parametric and nonparametric two-tailed unpaired t-tests, two-way analysis of variance (ANOVA), and Spearman's correlation between CEEVs and Fazekas scale were performed in Prism 10.0.2.We conducted a post hoc power analysis (G*Power 3.1) focused on the primary mechanism and our key area of interest-the NVC response in the left-dorsolateral prefrontal cortex (LDLPFC).This was to confirm that our sample size was robust enough to lend credibility to the study's conclusions.Upon averaging the data from the LDLPFC across groups, with 19 participants in the CN group and 18 in the MCI group, we reached a statistical power of 0.8072.This indicates a strong likelihood that our study was adequately powered to detect a true effect.Educational attainment was separated by bachelor's degree and higher, and other, to not violate chi-squared distribution assumptions.Brain AnalyzIR toolbox, implemented in MATLAB R2023b, was used for group-level analysis of NVC responses.Correlation matrices, data preprocessing, feature selection, model training, and validation were performed in Python 3.All data presented were shown as mean ± SD unless otherwise stated.Parametric tests were only used if the Shapiro-Wilk test showed normal distribution.Nonparametric Mann-Whitney test was used to compare task-averaged D, task-averaged LDLPFC NVC, CEEV%, CEEV concentration, EEV concentration, and MAL+ EV concentration.For two-way ANOVA statistical test, which was used for all measures assessing the effect across n-back tasks per group and size distribution of CEEVs, robust regression outlier removal (ROUT) method was used at the lowest tuning constant (Q = 0.1%) to remove extreme outliers that would violate homoscedasticity and normality of residuals across groups in parametric statistical tests.Bonferroni post hoc tests were used for the assessment of group accuracy, RT and FC during the n-back session, and CEEV size distribution.Spearman's correlation was used to assess the relationship between CEEVs and functional neuroimaging data and cognition (Jamovi 2.4.12.0).For NVC responses individual β weights for channel, subject and n-back task were obtained from first-level GLM.Mixed effects secondlevel GLM, fitted according to the Wilkinson-Rogers formula ('β ∼ −Group: session + (1|Subject)' , were evaluated by t-contrasts for each channel to compare CN and participants with MCI (unpaired two-tailed t-test), and between condition (e.g., 1b vs. 0b_1) within group (paired t-test).Type 1 error (from multiple [channel-wise] comparisons) was pants with MCI had historical MRI imaging.Participant MRIs were conducted within five years of participation in the study (median = 2022, IQR = 2021-2023).Figure 1 shows the representative image of one CN participant MRI (age: 68 years old, sex: female, Deep White Matter Fazekas scale = 0), and one participant with MCI (age: 65 years old, sex: female, Deep White Matter Fazekas scale = 1).All deep white matter Fazekas scale scores are shown in Supplementary Figure 3.
changes in reaction time (RT).One participant was removed from d' and RT 1-back (1b) and 2-back (2b) analysis in the MCI group and one participant was removed from 2b analysis in the CN group because they did not understand the task.An additional participant was removed from the CN group for all tasks because they clicked the wrong mouse button during the entire n-back task.There was a significant decrease in 2b performance (CN: n = 18, MCI: n = 19, 95% confidence interval (CI) of predicted mean difference = [0.4064 to 1.168], p < 0.0001) in MCI compared to CN.Additionally, 2b RT was increased in MCI (n = 19) compared to CN (n = 18, 95% CI of predicted mean difference = [−255.3to −26.54], p = 0.0088), indicating impaired working memory performance and processing speed during more cognitively challenging tasks.

F I G U R E 1
Clinical grading of small vessel ischemic lesions in control and participants with mild cognitive impairment (MCI).(A) Control (CN) participant (age: 68 years old, sex: female) with historical magnetic resonance imaging (MRI), recorded in 2023 shows a Fazekas scale grading of 0, indicating no small vessel ischemic damage determined by this grading classification.(B) Participant with MCI (age: 65 years old; sex: female) historical MRI, recorded in 2023, shows a Fazekas scale grading of 1, indicating moderate small vessel ischemic damage.Fazekas scale grading was assessed for deep white matter hyperintensities.F I G U R E 2 Fluid cognitive performance is diminished in participants with mild cognitive impairment (MCI), while crystalized intelligence is maintained in normal age rage.(A) Patients with MCI had a significant decrease in crystalized intelligence metric-picture vocabulary test (PIVT)-compared to control (CN), but maintained cognitive abilities in normal age, sex, race, and education range (Normative population mean = 50, SD = 10).Fluid cognition was impaired compared to age-matched CNs in domains of inhibitory control and attention (FICA), working memory (LSMT), cognitive flexibility (DCCS), episodic memory (PSMT), and processing speed (PCPS).Fully corrected T-score compares the test taker score to the nationally representative National Institutes of Health (NIH) toolbox sample, adjusting for age, sex, race, and educational attainment.(B) Processing speed was directly measured by oral symbol digit test and patients with MCI had decreased number of correct responses compared to age-matched controls.(C) Fluid cognition composite scores, including performance on FICA, LSMT, DCCS, PCPS, and PSMT, indicate overall fluid cognitive deficit in MCI compared to age-matched controls.Violin plots solid and dashed lines are presented as median and interquartile range, respectively; n = 20 participants per group.*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 by two-tailed Student's t-test (unpaired).

F I G U R E 3
Working memory performance and neurovascular coupling (NVC) is impaired in participants with mild cognitive impairment (MCI).(A) Accuracy (d') (see formula in main text) during 2b working memory n-back task was decreased in MCI compared to age-matched controls (CN) and (B) reaction time was increased during the most cognitively challenging 2b task.Panels C and D show the NVC response (relative change in oxy-hemoglobin [HbO]) during the n-back task.To determine the NVC response to cognitively challenging tasks, the contrast: (2-b+ 1-b) -0b_2 was used for CN, MCI, and (E) group comparison (MCI -CN).The black circle represents the left dorsolateral prefrontal cortex (LDLPFC), a region responsible for working memory.(C) CN group showed increased NVC (red shaded areas), and (D) MCI group showed no change to increased cognitive workload across the prefrontal cortex.(E) Group analysis of this contrast showed decreased NVC (blue-shaded areas) in the LDLPFC and medial prefrontal cortex in MCI compared to age-matched controls.(F) Averaged NVC responses across the entirety of the n-back paradigm determined that MCI patients had decreased NVC responses localized to the LDLPFC and medial PFC compared to controls.(G) Regression coefficients (beta) from (F) were extracted from the LDLPFC channels and show decreased NVC in MCI compared to age-matched controls.Violin plots (A, B) solid and dashed lines are presented as median and interquartile range and box and whiskers graph (G) is presented as median (solid line), interquartile range (IQR) and minimum and maximum.Panels C, D, E, and F t-contrast maps were generated using Brain AnalyzIR toolbox implemented pipeline based on General Linear Model approach.For further details see main text.For Panel A: 0b_1 n = 19 CN, 20 MCI, 0b_2 n = 15 CN, 16 MCI, 1b n = 15 CN, 19 CN, 2b n = 18 CN, 19 MCI.For Panel B: 0b_1 n = 19 CN, 20 MCI, 0b_2 n = 19 CN, 20 MCI, 1b n = 19 CN, 19 MCI, 2b n = 18 CN, 19 MCI.Panels A and B, robust regression outlier removal (ROUT) method was used for extreme outlier detection (tuning constant Q = 0.1%).**p < 0.01, ****p < 0.0001 by two-way analysis of variance (ANOVA) with Bonferroni's multiple comparisons test.For Panels C-G: n = 19 CN, 18 MCI.Panels C-F: Colors refer to t-statistics thresholded by q < 0.05 (obtained after false discovery rate [FDR] correction) Panel G: **p < 0.01 by unpaired t-test.0b_1, first 0-back; 0b_2, second 0-back; 1b, 1-back; 2b, 2-back; NVC, neurovascular coupling.and group comparison are presented in Supplementary Figure 4, which shows consistently decreased NVC in MCI compared to CN across nback conditions.Figure 2F shows a significant decrease in the LDLPFC NVC response across the entire n-back paradigm in MCI compared to controls.Regression coefficient (β) values extracted from general linear model analysis showed that MCI had significantly decreased averaged LDLPFC NVC (F3-F5, F3-F1, F3-FC3) responses (n = 18, median = 1.401,IQR = [−17.27to 14.16]) compared to CN (n = 19, median = 22.86, IQR = [9.180 to 37.30, p = 0.0033]) (Figure

F I G U R E 4
Figure 6B).
w D LDP , weighted task-averaged LDLPFC node degree; D, task-averaged global node degree; w D, weighted task-averaged global node degree.most informative variables for the final model.For the NVC variable (LDLPFC NVC), participant β values from LDLPFC channel-averaged NVC response across the duration of the n-back task were used.For FC, LDLPFC task-averaged w D LDP and D LDP were assessed to determine group classification.Additionally, known factors that have shown predictive value to classify MCI, such as age, sex, hypertension, 7 and depression prior to MCI diagnosis, 34 were variables in our feature selection process.Sample size for our group classification analysis consisted of all patients who understood, and successfully completed all tasks of the n-back task and had plasma samples available for CEEV analysis (CN n = 17, MCI n = 17).
Model evaluation metrics on the single test set with 500 decision trees highlighted the model's validity.It achieved an accuracy of 85.71%, precision of 88.57%, recall of 85.71%, and an F1 score of 85.08%.For mean accuracy estimation, the model underwent 100 runs with different test-train splits (80% train followed by 20% test) to ensure a comprehensive assessment of its performance.The aggregated metrics on the test set showed a mean accuracy of 68.0%, with an SD of 16.0%, indicating the reliability across diverse datasets, offering a more robust evaluation compared to a single test-train split.The integration of LOOCV, random Forest, and multiple test-train splits enhances the robustness of the model evaluation process, providing a comprehensive understanding of its performance on different datasets.The selected features and feature importance, along with the aggregated and individual metrics, collectively affirm the reliability and generalization capability of the developed classification model, determining new insights into the combined role of NVC, FC, and CEEVs in MCI etiopathology.
ipants with MCI compared to CN.These findings were localized to the LDLPFC.Exploratory study showed elevated ratio and concentration of CEEVs in MCI compared to CN. CEEVs significantly associated with increased WMH burden in participants with MCI.Following robust model and feature selection, NVC and FC in the LDLPFC, and CEEVs, were the most important variables in predicting MCI with high accuracy.These findings uncovered the alterations in cerebrovascular hemodynamics, and functional neuronal networks, while also showing CEEVs as a potential cerebrovascular pathologic marker that strongly associates with small vessel ischemic burden in participants with MCI.We confirmed fluid cognitive dysfunction in our MCI cohort by identifying overall fluid cognitive deficits, impairment in working memory, inhibitory control and attention, cognitive flexibility, and processing speed measured by n-back working memory test and NIH toolbox cognitive battery.The n-back test has long-standing validity as a working memory paradigm.10,38NIH toolbox cognitive battery has high discriminant ability that has found utility in clinical populations with cognitive disorders.39Through these multitest cognitive domain evaluations, we provided strong certainty of group differences (MCI vs. CN) in fluid cognitive functioning and preserved crystalized abilities.Moreover, we showed that CEEVs were sensitive to fluid cognitive performance in our participants, indicating the potential dual-role of CEEVs as biomarkers of cognitive performance and small vessel ischemic damage.

ACKNOWLEDGMENTSFigure 7 and
Figure 7 and Supplementary Figures 1 and 2A were created with Biorender.com.Research reported in this publication was supported by the National Institute on Aging of the National Institutes of Health under Award Number R01AG075834 and R21AG080775, and by the American Heart Association under Award Number 966924 and 23DIV-SUP1070991.This work was also in part supported by the National Institute on Aging (RF1AG072295, R01AG055395, R01AG068295, K01AG073614, K01AG073613, R03AG070479) the Eunice Kennedy Shriver National Institute of Child Health and Human Development (R01HD109157), the National Institute of Neurological Disorders and Stroke (R01NS100782, RF1NS085002), the U.S. Department of Veterans Affairs (Merit award number CX002578 and CX000340), the National Cancer Institute (R01CA255840) of the National Institutes of Health, the American Heart Association (AHA941290, AHA932980), the Oklahoma Shared Clinical and Translational Resources (OSCTR) program funded by the NIGMS (U54GM104938), and the NIA-funded Geroscience Training Program in Oklahoma (T32AG052363).The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
-distinguish CEEVs from EEVs.The ratios of cerebrovascular to vascular vesicles were assessed for each aliquot by dividing the number of endothelial-and MAL-positive vesicles by the total number of vesicles of endothelial origin (endothelial + /MAL + divided by endothelial + /MAL + + endothelial + /MAL − ) and expressed as percentage.Furthermore, overall CEEV concentrations were measured, averaged, and compared between groups.All concentration values were corrected by (multiplying by) dilution factor.For each subject, the average values from replicates were used.Representative images used in corresponding figures were created using FCS Express Pro 7 (v.7.18.0025,DeNovo Software).
Patient demographics, conditions, medications, and baseline physiology.
34± 8.0 years of age) and age-matched controls([CN]n = 20, 70.8 ± 6.6 years of age) are summarized in Table1.A significantly higher proportion of participants with MCI had ongoing, medically controlled psychiatric disease (n = 13) compared to CN (n = 5, TA B L E 1 plex virus, sleep apnea, asthma, chronic pancreatitis, dizziness, migraine, trigeminal neuralgia, carpal tunnel.Psychiatric disease: depression, anxiety, post-traumatic stress disorder.Other prescribed drugs: ezetimibe, acyclovir, tramadol, trazadone, bisphosphonates, celecoxib, metformin, chronic obstructive pulmonary disorder medication, omeprazole, apixaban, escitalopram, buspirone, antiviral hepatitis medication, clopidogrel, carbamazepine, bupropion, insulin, dapagliflozin, cyclobenzaprine, semaglutide, memantine, donepezil.p=0.0248), which includes depression, anxiety, and post-traumatic stress disorder.We evaluated the diagnostic date of psychiatric disease, specifically depression, as this made up the largest percentage of psychiatric disease in participants with MCI (85% of psychiatric disease cases).This distinction is critical, as there is a bidirectional relationship between MCI and depression.34Six of 11 participants with MCI were diagnosed with depression prior to MCI diagnosis.There was no significant difference in proportion of depression in CN (n = 2) compared to depression prior to MCI diagnosis (n = 6, p = 0.2351).Further, we investigated etiology of MCI based on data obtained from medical records and these data are presented in Supplementary Table1.One CN participant had a historical T2 FLAIR MRI and 17 partici- 10e disrupted in MCI and that these features provide high discriminate value compared to demographic and comorbidity factors.Moreover, these findings place a great deal of importance for early identification of NVC disruption, as this disruption predates the loss of cognitive compensation mechanisms (i.e., FC)10and MCI diagnosis.
51EV ratio, and LDLPFC FC had the highest feature importance in MCI classification, even with age, sex, depression, and hypertension data incorporated in feature selection.Previous studies have shown machine learning predicting MCI diagnosis based on patient reported cognitive decline49demographic and comorbidity variables,50and resting state FC.51Here, we determined that neurovascular health and FC