Comparing model‐based cerebrovascular physiomarkers with DTI biomarkers in MCI patients

Abstract Objective To compare the novel model‐based hemodynamic physiomarker of Dynamic Vasomotor Reactivity (DVR) with biomarkers based on Diffusion Tensor Imaging (DTI) and some widely used neurocognitive scores in terms of their ability to delineate patients with amnestic Mild Cognitive Impairment (MCI) from age‐matched cognitively normal controls. Materials & Methods The model‐based DVR and MRI‐based DTI markers were obtained from 36 patients with amnestic MCI and 16 age‐matched controls without cognitive impairment, for whom widely used neurocognitive scores were available. These markers and scores were subsequently compared in terms of statistical delineation between patients and controls. Results It was found that statistically significant delineation between MCI patients and controls was comparable for DVR or DTI markers (p < 0.01). The performance of both types of markers was consistent with the scores of some (but not all) widely used neurocognitive tests. Conclusion Since DTI offers a measure of cerebral white matter integrity, the results suggest that the model‐based hemodynamic marker of DVR may correlate with cognitive impairment due to white matter lesions. This finding is consistent with the hypothesis that dysregulation of cerebral microcirculation may be an early cause of cognitive impairment, which has been recently corroborated by several studies.


| INTRODUC TI ON
The quest for noninvasive diagnostic markers of early-stage Alzheimer's disease (AD) has led to the development of biomarkers derived from Magnetic Resonance Imaging (MRI) and Diffusion Tensor Imaging (DTI), in addition to the traditional neurocognitive testing scores. Individuals with amnestic Mild Cognitive Impairment (MCI) have high risk of developing AD, and MCI has been considered a transitional stage between normal aging and AD. Recently, modelbased "physiomarkers" of cerebral hemodynamics have also been introduced for the same purpose under the hypothesis that dysregulation of cerebral microcirculation may cause cognitive impairment (Marmarelis, Shin, Orme, & Zhang, 2013;Marmarelis, Shin, Tarumi, & Zhang, 2017). The use of MRI and DTI markers is based on the premise that the onset of neurodegenerative disease is associated with an elevated risk for structural and functional abnormalities in the brain that are detectable by MRI and/or DTI (Charlton et al., 2006;Madden, Bennett, & Song, 2009;Mori & Zhang, 2006;Sexton, Kalu, Filippini, Mackay, & Ebmeier, 2011;Smith et al., 2006;Sun et al., 2006;Tarumi et al., 2015;Wardlaw et al., 2013). Much attention has been accorded to the presence of white matter lesions as an important risk factor for cognitive impairment (Groot et al., 2013;Maillard et al., 2014;Young, Halliday, & Kril, 2008). Although markers based on structural MRI have some utility, more promise has been offered by DTI-based biomarkers that quantify the neuronal fiber integrity in white matter using measures such as Fractional Anisotropy (FA) and Radial or Mean Diffusivity (RD or MD) (Beaulieu & Allen, 1994;Mori & Zhang, 2006). Correlation between these DTI-based markers and neurocognitive performance has been demonstrated (Madden et al., 2009). Specifically, these DTI-based biomarkers have been shown to correlate with the myelination level and the density of the neuronal fibers (Song et al., 2003). Since the main cause of these white matter lesions may be rooted in abnormalities of cerebral blood circulation and cardiovascular regulation, leading to cerebral hypoperfusion or ischemia (Fernando et al., 2006;Moody, Bell, & Challa, 1990), we wish to explore possible correlations between these DTI-based biomarkers and model-based physiomarkers of cerebral hemodynamics that we have recently introduced (Marmarelis et al., 2017) in order to delineate MCI patients from age-matched controls without cognitive impairment. Our working hypothesis is that if white matter lesions are associated with abnormalities in cerebral microcirculation, then our cerebral hemodynamic physiomarkers may correlate with the DTIbased biomarkers of cerebral white matter neuronal fiber integrity.

| Study participants
This study enrolled 52 participants, 36 patients with amnestic MCI (19 female and 17 male), and 16 age-matched cognitively normal controls (eight female and eight male) through a community-based advertisement and the University of Texas Southwestern Medical Center Alzheimer's Disease Center. Inclusion criteria were as follows: men and women aged 55-80 years with normal cognitive function or MCI. Exclusion criteria included a history of cardiovascular (e.g., angina, myocardial infarction), cerebrovascular, or psychiatric disease, uncontrolled hypertension or dyslipidemia, diabetes mellitus, obesity (body mass index >35 kg/m 2 ), current or a history of smoking within the past 2 years, or chronic inflammatory disease. Individuals with a pacemaker or any metal in their body that precludes MRI were also excluded. The diagnosis of amnestic MCI was based on the Petersen criteria (Petersen et al., 1999), as modified by the Alzheimer's Disease Neuroimaging Initiative project. Clinical evaluation was based on the recommendations of the Alzheimer's Disease Cooperative Study (Morris et al., 1997). Specifically, MCI patients had to meet the following criteria: (a) subjective memory complaint, (b) a global Clinical Dementia Rating of 0.5 with a score of 0.5 in the memory category, (c) objective memory loss as indicated by

| Diffusion tensor imaging
A 3-T magnetic resonance imaging (MRI) scanner (Philips Medical System) was used to acquire the Diffusion Tensor Imaging (DTI) data using a single-shot echo planar imaging (EPI) sequence with a sensitivity encoding (SENSE) parallel imaging scheme (reduction factor = 2.2). The imaging matrix was 112 × 112 with field of view (FOV) = 224 × 224 mm 2 (nominal resolution of 2 mm), which was filled to 256 × 256 pixels. Axial slices of 2.2 mm thickness (no gap) were acquired parallel to the anterior-posterior commissure line. A total of 65 slices covered the entire hemisphere and brainstem. Echo time (TE)/repetition time (TR) was 51/5,630 ms. The diffusion weighing was encoded along 30 independent orientations, and the b value was 1,000 s/mm 2 . The scan duration was 4.3 min. Automated image registration was performed on the raw diffusion images to correct distortions caused by motion artifacts or eddy currents. The DTI scan was performed twice. For image preprocessing and voxel-wise tractbased spatial statistical (TBSS) analysis (Smith et al., 2004(Smith et al., , 2006, we used the broadly accepted software library of FMRIB (FSL, https :// surfer.nmr.mgh.harva rd.edu/fswiki). The ROI analysis was performed using the ICBM-DTI-81 white matter atlas (Mori et al., 2008).  (Drane, Yuspeh, Huthwaite, & Klingler, 2002;Wechsler, 1987). For episodic memory assessment, this study used the following three tests: California Verbal Learning Test with long delay free recall (CVLT-LDFR), California Verbal Learning Test total of parts 1-5 (CVLT-total) (Delis, Kramer, Kaplan, & Ober, 2000), and Visual Reproduction t-score recall (VRT).

| Model-based analysis of cerebral hemodynamic data
Cerebral hemodynamic time-series data were collected at the resting state over 5-6 min (after 20 min of rest) via transcranial Doppler (TCD) at an initial sampling rate of 1 KHz of blood flow velocity (CBFV) at the middle cerebral arteries using a 2 MHz TCD probe (Multiflow, DWL) placed over the temporal window and fixed at constant angle with a custom-made holder. Concurrently, continuous measurements of arterial blood pressure (ABP) were made with finger photo-plethysmography (Finapres) and of end-tidal CO2 (ETCO2) via a nasal cannula using capnography (Criticare Systems). All measurements are noninvasive, safe, and comfortable for the subjects.
The data were collected in a quiet, environmentally controlled laboratory under resting seated conditions. The raw data were reduced to beat-to-beat values following our established preprocessing procedures (Marmarelis et al., 2017). The resulting beat-to-beat time-series data were used to estimate predictive dynamic models following our novel modeling methodology (Marmarelis et al., 2017).
The obtained predictive input-output models were then used to compute the predicted CBFV response of each subject to a 5-s pulse input of ETCO2, while the other input (ABP) is kept at baseline. We have found in our previous work that the average of this model-predicted CBFV response over 5 s represents a differentiating "physiomarker" for MCI patients (relative to age-matched controls), which is termed "Dynamic Vasomotor Reactivity" (DVR) index, because it quantifies the cerebral flow response of each subject to a sudden (and short) change in ETCO2 (a surrogate for blood CO2 tension) (Marmarelis et al., 2017). We note that the DVR is distinct from conventional measures of cerebral vasoreactivity that are obtained with breath-holding or CO2 inhalation. Detailed on this modeling methodology and its application to cerebral hemodynamics can be found

| Statistical analysis
Our statistical analysis considered 11 different diagnostic markers or neurocognitive scores: the DVR, three DTI markers (FA, MD, and RD), four neurocognitive test scores for assessment of attention-executive function (Trails-B, WAIS-dig, WCST-per, LF-F&A&S), and three neurocognitive test scores for assessment of episodic memory (CVLT-LDFR, CVLT-total, and VRT), which are described above. The ability of each marker/score to differentiate between the MCI patients and the cognitively normal controls was first assessed with the t-statistic (p-value).
The Pearson pairwise correlations were also examined. Finally, a linear fixed effects statistical regression model was used to separate the effects of age, gender, and education (viewed as covariates) and subsequently examine the ability of the resulting markers/scores to differentiate between the MCI patients and the cognitively normal controls, as well as the possible linear pairwise relations that are revealed by linear regression lines. Statistical significance was set at p < 0.05.
Statistical analysis was performed using MATLAB (MathWorks Inc.). Table 1 shows demographics of control subjects (CS) and MCI patients (MP). Table 2 shows the mean (SD) values for each of the 11 selected markers/scores and the p-value for the statistical differentiation between the groups of 16 CS and 36 MP, before and after separating out the effects of the covariates (age, gender, and education). It is seen that 6 of the 11 markers/scores have p < 0.01 and one has p < 0.05. One DTI marker (FA) and three neurocognitive scores (WCST-per, LF-F&A&S, and VRT) do not allow statistically significant differentiation between MP and CS (p > 0.05) before separating the covariates effects, while only two neurocognitive scores (WCSTper and LF-F&A&S) do not allow statistically significant differentiation after separating the covariates effects.

| Pairwise correlations between the 11 markers/ scores
In order to examine possible pairwise correlations between the 11 markers/scores considered in this study, we show in Table 3 the Pearson correlation estimates between all pairwise combinations (after separating out the effects of the covariates of age, gender, and education) for all MP and CS taken together. The p-value of statistical significance of each pairwise correlation is also given in parentheses underneath the correlation estimate. We observe that the DVR physiomarker is only highly correlated with the three DTI biomarkers (marked in bold italics in Table 3 for p < 0.01) and weakly correlated with only one neurocognitive test score, viz. the CVLT-LDFR for episodic memory (0.01 < p < 0.05 marked in italics in Table 3). The latter also strongly correlates with the MD biomarker, while the related CVLT-total score correlates weakly with MD. As expected, the correlation between CVLT-LDFR and CVLT-total is strong. The three DTI biomarkers are

| Linear regression between DVR and the DTI biomarkers or the CVLT-LDFR score
Since the novel DVR physiomarker is the focus of this paper, its relations with the DTI biomarkers and neurocognitive scores are examined further in this section through regression analysis, which is limited to the markers/scores with which significant correlations exist (see Table 3). The scatter plots of DVR versus the three DTI biomarkers (FA, MD, and RD) extracted from the significant voxels and the CVLT-LDFR neurocognitive score of episodic memory are shown in Figure 1, along with the estimated regression lines. The r 2 values are substantial (and highly significant with p < 0.001) for DVR versus the three DTI biomarkers, but marginal for DVR versus the CVLT-LDFR score (r 2 = 0.09 and p = 0.029). We note that CVLT-LDFR is the memory-related neurocognitive score that shows the strongest correlation with a DTI biomarker, viz. MD (see Table 3).

| Associations of DVR with DTI biomarkers in specific brain regions
To provide more insight into the numbers of significant voxels of Fractional Anisotropy (FA), Mean Diffusivity (MD), and Radial Diffusivity (RD) that have been found in various brain regions by our analysis (i.e., the voxels where DTI biomarker values correlate significantly with DVR), we show in Table 4 those numbers of significant voxels and the respective percentages for 48 brain regions (R: right side, L: left side). The highest percentages (>30%) of significant voxels were found in the 22 brain regions highlighted in boldface in Table 4.  Table 2) after accounting for the effects of the covariates of age, gender, and education. We note that the efficacy of the DTI biomarkers is assessed over the "significant" voxels where significant correlation exists between the DTI biomarkers and the DVR physiomarker.

| D ISCUSS I ON
The main finding of this study is that significant pairwise Pearson correlations exist between the DVR physiomarker and the three DTI biomarkers of FA, MD, and RD (p < 0.001), as well as between the DTI biomarkers and the Trails-B neurocognitive test score (see Table 3). Furthermore, we found significant regression lines (associations) between DVR and the three DTI biomarkers, as well as the CVLT-LDFR neurocognitive score that is related to episodic memory (see Figure 1). We also found that only the MD DTI biomarker correlates significantly with the CVLT-LDFR and CVLT-total neurocognitive scores, and only four pairwise correlations of neurocognitive scores are significant (see Table 3). are physically restricted by axonal membranes and myelin (Mori & Zhang, 2006) and their values are thought to provide measures of the integrity of the WM neuronal fibers. Since the major WM fiber tracts that may be susceptible to cerebral hypoperfusion and/or ischemia are primarily found in the deep and periventricular brain regions, such as the corpus callosum, the corona radiata, the internal capsule, the external capsule, and the superior longitudinal fasciculus (Moody et al., 1990), it is notable that these brain regions are among those listed in Table 4 as having large percentages of significant voxels with high correlation between DTI biomarkers and the DVR physiomarker.

| Diagnostic equivalence between the DVR physiomarker and the DTI biomarkers
Cognitive function performance has been correlated with DTI biomarkers from global and regional WM fiber tracts in previous studies: processing speed and executive function (Charlton et al., 2006), processing speed and global cognition (Vernooij et al., 2009), and executive function (Tarumi et al., 2015)-while it was also shown that total brain volume of WM or WM hyperintensities were not correlated to cognitive performance. We note that, although the DTI biomarkers in global WM (quantifying the integrity of the WM neuronal fibers) were shown to correlate with executive function performance in MCI patients and cognitively normal controls alike, the MCI patients with lower executive function performance showed statistically similar levels of WM neuronal fiber integrity with cognitively normal controls (Tarumi et al., 2015). This intriguing discrepancy between the group statistical comparison and regression analysis may be due to the limited cohort size of these studies or to the multifactorial nature of cognitive impairment that includes amyloid deposition, hypometabolism, and cardiovascular/cerebrovascular dysregulation, among many others, which may influence cognitive function independently of defects in WM structural integrity (Arnaiz et al., 2001;Jack et al., 2009;Marmarelis et al., 2017;Tarumi et al., 2015). In the present study, when only the "significant" voxels of DTI brain images (i.e., those with significant correlation with DVR) were used for group statistical comparison, the DTI biomarkers delineated MCI patients from controls. The DVR physiomarker (being a global cerebrovascular measure for each subject) also achieves significant delineation of MCI patients from controls and exhibits significant pairwise Pearson correlations with the DTI biomarkers and the CVLT-LDFR score (see Table 3), as well as significant associations via regression analysis (see Figure 1). Note that the CVLT-LDFR test relates to episodic memory, while none of the neurocognitive tests related to executive function had significant pairwise correlation with DVR (see Table 3). This suggests that DVR may be related more to memory deficits than executive dysfunction.

| Results from specific brain regions
Using the significant voxels of DTI brain images (i.e., the voxels that show significant correlation between DVR and DTI values) over all subjects, we found that 22 of 48 analyzed brain regions contained more than 30% significant voxels (see Table 4). An illustrative example is shown in Figure 2 that demonstrates the localized correlations of DVR with DTI metrics in some brain regions. Among the 22 brain regions with high percentages (>30%) of significant voxels, we TA B L E 4 The numbers of significant voxels of Fractional Anisotropy (FA), Mean Diffusivity (MD), and Radial Diffusivity (RD)-that is, correlating significantly with DVR-which were found in 48 brain regions   [Diao et al., 2015]).
The findings of this study lead to the conclusion that the model-based cerebral hemodynamic index DVR, which constitutes a physiomarker capable of delineating amnestic MCI patients from cognitively normal controls (p < 0.01), correlates (p < 0.001) with the MRI-DTI biomarkers in various brain regions-especially with the ones perfused by the perforating arteries of the middle cerebral artery (which is the location of the TCD measurement of F I G U R E 2 Illustrative example of voxel-wise analysis of the correlation between DVR and localized DTI metrics in regions of brain white matter (computed from the significant voxels only): sagittal (left), axial (middle), and coronal (right) views, for FA (top panels), MD (middle panels), and RD (bottom panels). Color inside the brain maps indicates the areas that are significantly correlated with DVR (according to the color bar at the bottom of the Figure) Brain Note: The numbers in parentheses represent the percentage of significant voxels in each region.
TA B L E 4 (Continued) cerebral blood flow velocity from which the model-based DVR is derived), as well as with the scores of the CVLT-LDFR neurocognitive test that is related to episodic memory. These results corroborate the view that the DTI biomarkers may be influenced by dysfunction of brain microcirculation, and that DVR is a useful diagnostic physiomarker in conjunction with DTI biomarkers in patients with MCI.

ACK N OWLED G M ENTS
This work was supported in part by the NIH grants P41

CO N FLI C T O F I NTE R E S T
The authors have no conflict of interest to report.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from the corresponding author upon reasonable request.