Altered Hierarchical Gradients of Intrinsic Neural Timescales in Mild Cognitive Impairment and Alzheimer's Disease

Abstract

Alzheimer's disease (AD) is a devastating neurodegenerative disease that affects millions of seniors in the United States. Resting-state functional magnetic resonance imaging (rs-fMRI) is widely used to study neurophysiology in AD and its prodromal condition, mild cognitive impairment (MCI). The intrinsic neural timescale (INT), which can be estimated through the magnitude of the autocorrelation of neural signals from rs-fMRI, is thought to quantify the duration that neural information is stored in a local circuit. Such heterogeneity of the timescales forms a basis of the brain functional hierarchy and captures an aspect of circuit dynamics relevant to excitation/inhibition balance, which is broadly relevant for cognitive functions. Given that, we applied rs-fMRI to test whether distinct changes of INT at different hierarchies are present in people with MCI, those progressing to AD (called Converter), and AD patients of both sexes. Linear mixed-effect model was implemented to detect altered hierarchical gradients across populations followed by pairwise comparisons to identify regional differences. High similarities between AD and Converter were observed. Specifically, the inferior temporal, caudate, and pallidum areas exhibit significant alterations in both AD and Converter. Distinct INT-related pathological changes in MCI and AD were found. For AD/Converter, neural information is stored for a longer time in lower hierarchical areas, while higher levels of hierarchy seem to be preferentially impaired in MCI leading to a less pronounced hierarchical gradient. These results inform that the INT holds great potential as an additional measure for AD prediction, even a stable biomarker for clinical diagnosis.

Significance Statement

We observed high similarities of intrinsic neural timescales (INTs) between patients with Alzheimer's disease (AD) and people that will later progress to AD (called Converter), deviating from cognitively normal (CN) individuals. This indicates that the pathological excitation/inhibition imbalance already started before the conversion to AD. We also revealed distinct pathophysiological changes in stable mild cognitive impairment (MCI) and AD/Converter. For the AD and Converter, neural information is stored for a longer time in lower brain hierarchical areas, while higher levels of the hierarchy seem to be preferentially impaired in stable MCI. These results suggest the potential for INT as an additional measure for AD prediction, even a stable biomarker for clinical diagnosis.

Introduction

Alzheimer’s disease (AD) is a devastating neurodegenerative disease, beginning with mild memory loss and progressively leading to compromises in other cognitive domains and impairments in everyday function (Knopman et al., 2021). Resting-state functional magnetic resonance imaging (rs-fMRI) has been widely used to study neurophysiology in AD and its prodromal condition, mild cognitive impairment (MCI), because of magnetic resonance imaging’s (MRI's) relative simplicity of use, noninvasive nature, and relatively high spatial resolution.

The rs-fMRI functional connectivity (FC) in brain networks, which measures inter-regional synchrony detected from the blood oxygenation level-dependent (BOLD) fMRI sequence, is an emerging AD biomarker that holds promise for early diagnosis (Badhwar et al., 2017; Ibrahim et al., 2021). Multiple studies have reported reduced connectivity in the default mode network (Greicius et al., 2004; Grieder et al., 2018)—a network that includes lateral and medial prefrontal lobe, lateral temporal lobe, and superior parietal lobe that is considered to be a predictor of cognitive decline (Koch et al., 2012). The salience network, a network involving insula and anterior cingulate regions, is thought to be involved in the selection of goal-directed behavior in patients with AD and MCI (Binnewijzend et al., 2012). However, FC cannot reflect the fundamental organizational principles that characterize the functional architecture of the human brain (Jonas and Kording, 2017; Stocco et al., 2021).

In contrast, the intrinsic neural timescale (INT) is thought to quantify the duration that neural information is stored in a local cortical circuit—a property that directly supports functional specialization (Watanabe et al., 2019). Such neural timescales demonstrate a distinct gradient in the brain (Kiebel et al., 2008), where densely interconnected associative regions, such as prefrontal and parietal areas, have longer timescales compared with primary sensory areas. This heterogeneity of timescales is considered a basis of the brain functional hierarchy. INT can be estimated from the magnitude of temporal autocorrelations in rs-fMRI signals (Hasson et al., 2015). It has been shown that metrics using temporal autocorrelations of rs-fMRI are reliable, causally linked to biological processes beyond an FC network framework, and relevant to neuropsychiatric disease including dementias (Shinn et al., 2023).

INT captures an aspect of circuit dynamics relevant to excitation/inhibition (E/I) balance, which is broadly relevant to cognitive functions (Li and Wang, 2022). Specifically, large-scale biophysical modeling has demonstrated that INT is dependent on the strength of recurrent excitation—a biophysical property shown to support working memory and perceptual integration (Cavanagh et al., 2020). Furthermore, several disorders with cognitive dysfunction [e.g., schizophrenia (Wengler et al., 2020), Parkinson's disease (Wei et al., 2023), and epilepsy (Xie et al., 2023)] have shown alterations in INT, and this may be relevant to other neuropsychiatric disorders including AD. One of the pathological hallmarks of AD is the accumulation of amyloid-β (Aβ) peptides in the brain that occurs long before the clinical onset; synaptic failure due to axonal terminal dysfunction (reduced action potentials) when it is adjacent to amyloid neuritic plaques alters the E/I balance (Yuan et al., 2022). Studies in mouse models suggest that network hyperexcitability can occur in early stages of AD and contribute to cognitive decline (Vossel et al., 2017). Given that the pathological distribution of Aβ plaques and tau paired helical filaments generally spares primary somatosensory and motor cortices and primary visual cortex until very late in the disease process while compromising multiple association cortices, the normal hierarchical gradient of INT could potentially be altered (Conejero-Goldberg et al., 2011; Serrano-Pozo et al., 2011).

Given the extensive synaptic loss and neuronal death in medial temporal regions (i.e., entorhinal and hippocampus) and neocortical regions in the AD dementia phase, we hypothesize that there will be widespread reductions in INT, that is, a shortening of the duration in which signaling is maintained in local circuits. To investigate, we used rs-fMRI to examine changes in INT across neural hierarchies and their hierarchical gradients in individuals with MCI, those progressing to AD, and AD patients.

Materials and Methods

Participants

Data used in this study were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). ADNI was launched in 2003 as a public–private partnership. The primary goal of ADNI has been to test whether serial MRI, positron emission tomography, other biological markers, and clinical and neuropsychological assessments can be combined to measure the progression of MCI and early AD.

A detailed description of the ADNI cohort has been previously published (Petersen et al., 2010). ADNI has recruited individuals that are CN, have MCI, or have dementia or AD. We used the rs-fMRI data of 945 subjects at the baseline. We excluded participants with poor data quality (see below, Image acquisition and preprocessing, for details), and the final sample included 904 participants. Furthermore, we divided the subjects into four groups based on their baseline and end-visit clinical status (up to 10 years or until they drop off the study): (1) CN group if the subject was CN at the baseline and not demented at end visit; (2) MCI group if the subject was MCI at the baseline and not demented at end visit; (3) Converter group if the subject was CN/MCI at the baseline and demented at end visit; and (4) AD group if the subject was demented at the baseline. Demographic characteristics are shown in Table 1. Of the CN/MCI individuals who progressed to AD (called Converters hereafter), four were CN at the baseline, while the remainder were diagnosed as MCI at the baseline.

Table 1.

Age (in years) and sex distribution by the baseline and end-visit clinical diagnosis group of the ADNI subjects

	CN (N = 476)	MCI (N = 262)	Converter (N = 57)	AD (N = 109)	Total (N = 904)	p value
Age						0.006
Mean	73.2	73.5	75	75.9	73.7
Min	55.7	55.6	57.3	55.3	55.3
Q1, Q3	67.7, 77.9	68.7, 79.1	70.2, 79.3	71.3, 81.8	68.4, 79.0
Max	95.5	93.3	89.7	96	96
Sex						0.002
F	281 (59.0%)	122 (46.6%)	24 (42.1%)	54 (49.5%)	481 (53.2%)
M	195 (41.0%)	33 (57.9%)	140 (53.4%)	55 (50.5%)	423 (46.8%)

Extended data of their APOE status, amyloid positivity, and the time to event for AD diagnosis is included in Extended Data Table 1-1. Age and sex distribution based on baseline diagnosis is included in Extended Data Table 1-2.

Table 1-1

Extended data for Table 1, presenting age (in years) and sex distribution, related AD biomarkers including APOE4 status, amyloid positivity extracted from PET imaging and hippocampal volumes, and the time to event (TTE) for AD diagnosis (measured in days; for CN and MCI groups, this was the censored time). The group categorization is based on the baseline and end-visit status of the ADNI subjects. N-Miss: number of missing, N: negative, P: positive; Q1: the first quartile, Q3: the third quartile. HC: hippocampal. Download Table 1-1, XLSX file ^{(11.4KB, xlsx)} .

Table 1-2

Extended data of Table 1 showing age and sex distribution based on baseline status of the ADNI subjects. Download Table 1-2, XLSX file ^{(9.8KB, xlsx)} .

Image acquisition and preprocessing

All MRI data were downloaded on August 18, 2020. T1-weighted (T1w) MRI and rs-fMRI data from 3 T MRI scanners were included in this study.

T1w images were acquired with a gradient-recalled–echo pulse sequence with acquisition in the sagittal plane with the following acquisition parameters: repetition time (TR), 6.98 ms; echo time (TE), 2.85 ms; inversion time (TI), 400 ms; field of view, 26 cm; voxel size, 1.0 × 1.0 × 1.2 mm; and acquisition matrix, 256 × 256 × 196. All T1w images underwent automated quality control through MRIQC (Esteban et al., 2017). For subjects with multiple available T1w images at the same visit, we selected the images with the best quality for further analysis (i.e., only one T1w image was selected for each subject). For subjects with multiple available T1w images at the same visit, we selected the images with the best quality for further analysis (i.e., only one T1w image was selected for each subject). For all selected T1w images that passed the quality check, cross-sectional image processing was performed using FreeSurfer Version 7.1.1 (https://surfer.nmr.mgh.harvard.edu/). Region of interest (ROI)-specific cortical thickness (CT) values were extracted from the automated anatomical parcellation using the Desikan–Killiany atlas (Desikan et al., 2006). These CT values were used as covariates to examine the effect of atrophy, since cortical thinning has always been observed in neurodegenerative disorder (Du et al., 2007) and it might potentially affect the INT values.

Rs-fMRI data were acquired with an echoplanar imaging sequence with the following acquisition parameters: 140 functional volumes; TR, 3,000 ms; TE, 30 ms; flip angle, 80; number of slices, 48; slice thickness, 3.3 mm; voxel size, 3 × 3 × 3.3 mm; and in-plane matrix, 64 × 64. For each subject, the first 5 volumes of the functional images were discarded for signal equilibrium and to allow the participant's adaptation to the scanning circumstances, leaving 135 resting-state volumes for further preprocessing. The preprocessing steps for rs-fMRI are described as follows. First, we used a custom methodology of fMRIPrep to generate a reference volume and its skull-stripped version. Head-motion parameters with respect to the BOLD reference including transformation matrices and six corresponding rotation and translation parameters are estimated before any spatiotemporal filtering using MCFLIRT under FSL 5.0.9 (Jenkinson et al., 2002). We then used 3dTshift from AFNI 20160207 for slice-time correction. The BOLD reference was coregistered to the T1w reference using bbregister by FreeSurfer, which implemented boundary-based registration with six degrees of freedom (Greve and Fischl, 2009). We resampled the BOLD time series into standard MNI152NLin2009cAsym space. Several confounding time-series metrics were calculated based on the preprocessed BOLD, that is, the framewise displacement (FD; Power et al., 2012) and the root-mean-square difference (RMSD) of the time series in the consecutive volumes (Jenkinson et al., 2002). Contaminated volumes were then detected and classified as outliers by the criteria FD > 0.5 mm or RMSD > 0.3% and replaced with new volumes generated by the linear interpolation of adjacent volumes using the CONN toolbox. The three global signals are extracted within the cerebrospinal fluid (CSF), the white matter (WM), and the whole-brain masks. We further bandpass filtered the time series with cutoff frequencies of 0.01 and 0.09 Hz. Finally, the covariates corresponding to head motion (six realignment parameters), outliers, and the BOLD time series from the subject-specific WM and CSF masks were removed from the BOLD functional time series using linear regression. After estimating voxel-wise INT values (see below, INT calculation) from the time series, ROI-specific INT were calculated as the mean of the INT values of the voxels within the ROI from the automated anatomical parcellation using the Desikan–Killiany atlas (Desikan et al., 2006) for cortical areas, which results in 68 cortical ROIs (34 on each hemisphere), and the Aseg atlas (Fischl et al., 2002) for subcortical areas. The Desikan–Killiany atlas has been widely used in previous AD studies since other refined brain parcellations do not fit well with aging brain, especially those with neurodegenerative disorder (Du et al., 2007; Tustison et al., 2014), and it facilitates the comparison with (and potential confound of) CT. We selected 16 subcortical ROIs (8 on each hemisphere), which includes the thalamus, caudate, putamen, pallidum, accumbens area, hippocampus, amygdala, and ventral diencephalon.

We excluded the participants (1) whose rs-fMRI contained >30% frames flagged as motion outlier (27 subjects were excluded), (2) without diagnosis information (12 were excluded), or (3) who had corrupted MRI images (two were excluded). Therefore, 904 participants (age range, 55–96, 481 females) were left for the analysis.

INT calculation

Before estimating the voxel-wise INT values, preprocessed rs-fMRI data were further processed with the following steps: (1) motion censoring of outliers that are consecutive in three and above volumes to be excluded and those volumes will be treated as NA and (2) spatial smoothing with a 6 mm full-width at half-maximum Gaussian kernel. For each subject $i$ and voxel $v,$the autocorrelation function (ACF) was estimated by the following:

$${\mathrm{ACF}}_{iv}^k={\displaystyle \frac{{\sum }_{t=k+1}^T(y_{iv,t}-{\overline{y}}_{iv})(y_{iv,t-k}-{\overline{y}}_{iv})}{{\sum }_{t=1}^T{(y_{iv,t}-{\overline{y}}_{iv})}^2},(1)}$$ (1)

where $k$ is the time lag, $k=1,2,\dots ,T-1,$ $T$ is the total number of timepoints, and ${\mathit{y}}_{iv}=(y_{iv}^1,y_{iv}^2,\dots ,y_{iv}^T)$ is the rs-fMRI signal sequence of voxel $v$ for subject $i.$ INT was then estimated as the area under the curve of the ACF during the initial positive period:

$$\mathrm{IN}{\mathrm{T}}_{iv}=\mathrm{TR}\sum _{k=1}^{N_{iv}}{\mathrm{ACF}}_{iv}^k,(2)$$ (2)

where $\mathrm{TR}$ is the repetition time and $N_{iv}$ is the lag directly proceeding the first negative $\mathrm{ACF}$ value of voxel $v$ for subject $i.$ Finally, the ROI-specific INT, $\mathrm{IN}{\mathrm{T}}_{ij},$ was calculated as the mean of the $\mathrm{IN}{\mathrm{T}}_{iv}$'s values within ROI $j.$

Scanner harmonization to remove site effect

The ADNI data were collected from 58 sites. To remove undesired artifacts from scanner and site differences, we applied the ComBat harmonization method (Fortin et al., 2017) to the INT and CT values. For each subject $i,$ ROI $j,$ and site $k,$ we fitted the following model:

$$\mathrm{IN}{\mathrm{T}}_{ijk}={\alpha }_j+{\gamma }_{jk}+{\mathit{X}}_i^T{\beta }_j+{\eta }_{ij}+{\delta }_{jk}{\epsilon }_{ijk},(3)$$ (3)

where ${\mathit{X}}_i^T$ contains the covariates of interest, which includes age, sex, mean framewise displacement (MFD), and diagnosis group. The ComBat harmonized INT was as follows:

$${\mathrm{INT}}_{ijk}^{\mathrm{combat}}={\displaystyle \frac{\mathrm{IN}{\mathrm{T}}_{ijk}-{\hat{\alpha }}_j-{\hat{\gamma }}_{jk}-{\mathit{X}}_i^T{\hat{\beta }}_j-{\hat{\eta }}_{ij}}{{\hat{\delta }}_{jk}}+{\hat{\alpha }}_j+{\mathit{X}}_i^T{\hat{\beta }}_j+{\hat{\eta }}_{ij}.(4)}$$ (4)

Similarly, we obtained the ComBat harmonized CT ${\mathrm{CT}}_{ijk}^{\mathrm{combat}}.$

Linear mixed-effect model to detect altered hierarchical gradient effects on INT

Our previous study has validated INT as an index of cortical hierarchy using the Glasser MMP1.0 atlas (Wengler et al., 2020). Here, to facilitate the comparison with previous literature and with CT alterations, we defined the hierarchical level (HL) of the 68 ROIs by the Desikan–Killiany atlas using the rs-fMRI of 100 unrelated young and healthy subjects from the Human Connectome Project (HCP) WU-Minn Consortium (Van Essen et al., 2013). Subcortical ROIs were excluded from hierarchical analyses given the disparate factors that shape the cortical hierarchy compared with subcortical hierarchies [where multiple hierarchies may exist within individual subcortical regions (Raut et al., 2020)]. We used the average INT map in 2 mm isotropic MNI space from 100 unrelated HCP young-adult subjects as in Wengler et al. (2020). The ROI-specific INT is calculated as the mean INT within each ROI. The HL is determined by ranking each ROI's INT value from low to high corresponding to short to long INT and unified to [0, 1] by dividing the total number of ROIs. The HLs of Desikan–Killiany atlas are visualized in Figure 1A, and the exact hierarchical ordering is given in Extended Data Figure 1-1. We compared the HLs between the Glasser and Desikan–Killiany atlases and confirmed that the Desikan–Killiany atlas maintains the HLs as the Glasser atlas.

Figure 1.

INT as a measure of brain HL. A, Comparison of the HLs of the Glasser and Desikan–Killiany atlases from the HCP dataset. Extended data table of the hierarchical orderings corresponding to the Desikan–Killiany atlas is included in Extended Data Figure 1-1. B, Parcellated group-averaged INT maps by groups from the ADNI database. The flame scale is ordered so that long (high) INTs are toward red and short (low) INTs are toward black.

Figure 1-1

Extended data table to Figure 1 A showing hierarchical orderings of the 68 ROIs by Desikan-Killiany Atlas. Download Figure 1-1, XLSX file ^{(13.4KB, xlsx)} .

We applied the linear mixed-effect (LME) model to predict the INT value in our ADNI sample using HL in the cerebral cortex, assuming different intercepts and slopes by diagnosis group. We also considered age, sex, MFD, and CT as covariates (fixed effects) and allowed for variations of the intercept and slope at the subject level (random effects). As such, for each subject i, and ROI j:

$$\begin{array}{rl}{\mathrm{INT}}_{ijk}^{\mathrm{combat}}=& a+b_h\mathrm{H}{\mathrm{L}}_j+{\mathit{X}}_{g,i}^T{\mathit{b}}_j+({\mathit{X}}_{g,i}^T*\mathrm{H}{\mathrm{L}}_j){\mathit{b}}_g^s+b_s\mathrm{Se}{\mathrm{x}}_i+b_a\mathrm{Ag}{\mathrm{e}}_i\\ & +b_m\mathrm{MF}{\mathrm{D}}_i+b_c{\mathrm{CT}}_{ij}^{\mathrm{combat}}+b_i+b_i\mathrm{H}{\mathrm{L}}_j+{\epsilon }_{ij},\end{array}(5)$$ (5)

where $\mathrm{Se}{\mathrm{x}}_i=I(\mathrm{Sex}=\mathrm{male})$, ${\mathit{X}}_{g,i}^T=(I(\mathrm{grou}{\mathrm{p}}_i=\mathrm{MCI}),I(\mathrm{grou}{\mathrm{p}}_i=$ $\mathrm{Converter}),I(\mathrm{group}\mathrm{\_}i=\mathrm{AD}))$ is a vector that contains the diagnosis group information for subject $i.$ T statistics and the corresponding p value were calculated for the regression coefficients, and their effect size was measured by partial ${\eta }^2,$ denoted as ${\eta }_p^2$ (Lakens, 2013).

Pairwise comparison to detect significant INT differences among groups

We conducted a pairwise comparison between diagnosis groups to identify significant INT differences. To streamline our analysis and reduce the number of comparisons, we averaged the INT values of the same ROIs on the left and right hemisphere, which left 34 cortical ROIs and 8 subcortical ROIs for comparison. The decision was further justified by the following: (1) the absence of clear laterization in AD, as shown in Extended Data Table 1-1, the hippocampal volume between the left and right hemispheres has not shown significant asymmetry across groups; and (2) symmetric INT patterns, the hierarchical gradient of INT values, as depicted in Extended Data Figure 1-1, appeared to be symmetric. Additionally, prior research (Wengler et al., 2020) has demonstrated symmetric relationships between INT and cognition/behavior.

The hypothesis testing was implemented through the following steps: first, ANOVA analysis was applied to select the ROIs that have group differences in INT using F tests at a significance level $\alpha =0.05$ with FDR correction; we included age, gender, MFD, and CT as covariates. In this step, we addressed the multiple comparisons by ROIs. Subsequently, among these selected ROIs, pairwise t tests were conducted to investigate where the differences presented between the four groups. Multiple-comparison correction was performed using the Tukey–Cramer method for comparing a family of four estimates, but not controlled for the 34 ROI comparisons since these pairwise comparisons are post hoc analyses. The effect size of the pairwise difference was measured by Cohen's d.

Sensitivity analysis

To test the robustness of the INT by HL, we did a series of sensitivity analysis and compared the results using Equation 5. Specifically, we repeat the LME model (1) without removing site effects using ComBat for INT and CT, considering the site is a potential source of artifacts in the neuroimaging data; (2) removing age from the covariates, since age and diagnosis are highly correlated which may attenuate the group effect; (3) removing CT from the covariates to examine the effect of atrophy; (4) including APOE4 positivity (which includes e4/e4, e3/e4, and e2/e4 genotypes)—a strong risk factor gene for AD—as a covariate; (5) including total hippocampal volume as a covariate; (6) replacing the four groups with three groups (CN, MCI, and AD) only based on their baseline diagnosis; and (7) excluding the four low-ranked ROIs (entorhinal and parahippocampus) to avoid low HL areas implicated in AD pathology overweighing the differences. Additionally, head motion is a common source of artifacts in fMRI data (Power et al., 2012); we included MFD as a covariate in all analyses and investigate its effect on the results.

In the current dataset, 114 participants (12.6%) did not have APOE4 information; thus a subset of $n=790$ subjects was included for this analysis. Details can be found in Extended Data Table 1-1. The demographic information of the groups based on baseline diagnosis can be found in Extended Data Table 1-2.

Code and data availability

R code is available for INT calculation at the GitHub website (https://github.com/Aiying0512/INT). Specifically, we used the acf function which allows for NA/missing values. ComBat harmonization was implemented using the R package neuroCombat. LME models were implemented using the R package lmerTest. The pairwise comparison of multiple groups was implemented using the R package emmeans.

ADNI datasets are available to the research community upon request at www.adni.loni.usc.edu. The processed imaging data are available for the qualified investigators upon request at seonjoo.lee@nyspi.columbia.edu.

Results

Hierarchical gradient effect on cortical INTs among various groups

Figure 1B shows the parcellated INT maps on the cerebral cortex for each of the four diagnostic groups and the 100 unrelated HCP subjects used to determine the cortical HLs along with the corresponding map of cortical HL. Following previous work (Wengler et al., 2020), our main analysis focused on hierarchical gradient effects on INT (i.e., the relationship between INT values and HL). The parameters of fitted lines of INT values as a function of HLs of the four diagnosis groups are visualized in Figure 2A. The detailed results of the LME model are shown in Extended Data Figure 2-1. This analysis only included the cortex given prior work demonstrating a separate hierarchical system in subcortical regions (Raut et al., 2020).

Figure 2.

The results of the LME fitting. A, The estimated intercept and slope of the hierarchical gradient effects by diagnosis groups and the corresponding 95% confidence intervals. Data corresponding to this figure are included in Extended Data Figure 2-1. B, Pairwise comparisons of the hierarchical gradient effects between the groups with significant differences. INT values were plotted as a function of HL, as well as their 95% confidence intervals. The Tukey–Cramer method was applied to adjust for comparing a family of four estimates. Data corresponding to this figure are included in Extended Data Figure 2-2.

Figure 2-1

Extended data table for Figure 2 A showing the estimated parameters in the linear mixed effect model. CI: confidence interval; ES: effect size; df: degree of freedom; HL: hierarchical level; MFD: mean framewise displacement. Download Figure 2-1, XLSX file ^{(13.1KB, xlsx)} .

Figure 2-2

Extended data table for Figure 2 B showing pairwise comparison results of hierarchical gradient effects among diagnosis groups. ES: effect size, which is measured by Cohen's d statistics; df: degree of freedom Download Figure 2-2, XLSX file ^{(11.6KB, xlsx)} .

In terms of differences between diagnostic groups in their relationship between INT values and HL (Fig. 2B), the MCI group had a less pronounced hierarchical gradient effect (as reflected in the contrasts of slope) when compared with the CN group in which INT values were similar in lower-order areas but were shorter in higher-order areas; the AD group showed longer INT values in lower-order areas than the CN group as reflected by the contrasts of the intercept; the AD and Converter groups had longer INT across all cortical areas compared with the MCI group. The full pairwise-comparison results are shown in Extended Data Figure 2-2.

Assessment of robustness of the INT by hierarchical gradient effect

We set out to determine the robustness of differences between diagnostic groups in their relationship between INT values and HL through a series of sensitivity analyses. The p values for linear mixed models are based on the degree of freedom with Satterthwaite approximation. As shown in Figure 3, the basic findings of INT by HL under various settings of sensitivity analyses are consistent with the main results in Figure 2, showing robustness to ComBat harmonization, alternative diagnostic group definitions, not covarying for age and CT, covarying for APOE4 status and total hippocampal volume, and exclusion of the four lowest HL cortical areas.

Figure 3.

The LME results of the sensitivity analyses: the estimated (A) intercept and (B) slope of the hierarchical gradient effects by diagnosis groups and the corresponding 95% confidence intervals are presented. The results of higher intercepts in AD/Converter group and the lower slope in the MCI group maintain across various sensitivity analyses. Inc APOE4, include APOE4 as a covariate; BL DX, use baseline diagnosis; Exc Age, exclude age as a covariate; Exc CT, exclude CT as a covariate; Exc Low-HL, exclude low HL regions; Inc HCV, include hippocampal volume as a covariate, W/o Combat, without applying ComBat harmonization. Data corresponding to this figure are included in Extended Data Figure 3-1.

Figure 3-1

Extended data table for Figure 3 showing the estimated parameters in the linear mixed effect models under various sensitivity analysis settings. CI: confidence interval; ES: effect size; df: degree of freedom. Download Figure 3-1, XLSX file ^{(20.9KB, xlsx)} .

No significant effects of head motion (quantified as MFD) were observed in the main results (Extended Data Figure 2-1) or any of the sensitivity analyses mentioned above (all p values >0.4 in Extended Data Figure 3-1). A significant age effect was observed in the main results (p value = 0.019): INT was shorter as age increased when controlling for other covariates. When we excluded the age to fit the LME model, the positive effect of the AD group on INT was no longer significant (p value = 0.055 in Extended Data Figure 3-1), likely due to the close relationship between AD and age. When total hippocampal volume is included as a covariate, the differences in intercepts between AD and CN (p value = 0.01) and Converter and CN (p-value = 0.026) are both significant, further highlighting their similarity. We compared the main results and the ones without applying ComBat to remove site effects (Fig. 3, Extended Data Figure 3-1) and found site effects reduced the significance of age influence, suggesting that the removal of site effects reduced nonbiological variability in the neuroimaging data while maintaining meaningful biological variability. No significant CT effect was observed for hierarchical gradients of INT (although see below for effects in individual ROIs).

Significant cortical ROIs in which AD and Converter groups had longer INT values than CN and MCI groups

Even though excluding low-ranked ROIs (entorhinal cortex and parahippocampus) slightly attenuated the differences between diagnostic groups in their relationship between INT values and HL (Fig. 3), the basic findings were still present. Therefore, it is critical to understand what ROIs drive the differences in these relationships. To this end, we conducted pairwise comparisons to examine the 34 cortical ROIs (averaged across the hemisphere).

Although CT had no significant effect on the relationship between INT and HL, we included the ROI's CT (averaged across the hemisphere) as one of the covariates to adjust the partial (local) volume effect on INT, thereby controlling for regional atrophy. After FDR correction, none of the regional CT effects on INT were significant. Additionally, sex, age, and MFD were adjusted for in the pairwise comparisons.

As shown in Figure 4, three ROIs were identified as having higher INT values in the AD group than in the CN group: entorhinal (CN–AD, $\mathrm{estimate},-0.086$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.550$; $t_{(895)}=-4.401$; $p=0.0001$), fusiform (CN–AD, estimate, $-0.044$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.298$; $t_{(896)}=-2.612$; $p=0.0452$), and inferior temporal (CN–AD, $\mathrm{estimate},-0.050$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.348$; $t_{(896)}=-3.090$; $p=0.0111$). The full comparison results can be found in Extended Data Figure 4-1.

Figure 4.

Cortical ROIs in which the INT values in the AD group (and the Converter group) are significantly longer than those in CN and MCI groups. A, Brain visualization of the ROIs. The colormap indicates the significance of group differences in INT by ANOVA test (p values and q values from FDR correction are provided). B, The boxplots of the INT values by diagnosis groups on the identified ROIs. For each ROI, the mean INT values are shown by group on the bottom of the boxplot, and the significant pairwise p values from post hoc comparisons are given on the top (threshold at 0.05). Age, sex, MFD, and CT were adjusted. The Tukey–Cramer method was applied to adjust for comparing a family of four estimates. Data corresponding to the comparison results are included in Extended Data Figure 4-1.

Figure 4-1

Extended Data for Figure 4. Pairwise comparison results of the cortical ROIs in which the INT values in AD group (and Converter group) are significantly longer than those in CN and MCI groups. ES: effect size, which is measured by Cohen's d statistics; df: degree of freedom. The unadjusted p value, Tukey-Cramer adjusted p value and the q value by FDR correction are provided. Download Figure 4-1, XLSX file ^{(11.8KB, xlsx)} .

Significant cortical ROIs in which the MCI group had shorter INT values than other groups

As shown in Figure 5, five ROIs were identified as having significantly shorter INT values in the MCI group than in the AD group, which includes caudal middle frontal (MCI–AD, $\mathrm{estimate},-0.068$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.193$; $t_{(896)}=-2.891$; $p=0.0205$), lingual (MCI–AD, $\mathrm{estimate},-0.078$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.219$; $t_{(896)}=-3.283$; $p=0.0059$), middle temporal (MCI–AD, $\mathrm{estimate},-0.065$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.200$; $t_{(896)}=-3.002$; $p=0.0146$), precentral (CN–MCI, estimate, $0.036$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=0.174$; $t_{(896)}=2.601$; $p=0.0465$), and superior temporal (MCI–AD, $\mathrm{estimate},-0.068$; Cohen's $d=-0.215$; $t_{(896)}=-3.217$; $p=0.0073$) areas. The full comparison results can be found in Extended Data Figure 5-1.

Figure 5.

Cortical ROIs in which the INT values of the MCI group are significantly shorter than other groups. A, Brain visualization of the ROIs. The colormap indicate the significance of group differences in INT by ANOVA test (p values and q values from FDR correction are provided). B, The boxplots of the INT values by diagnosis groups on the identified ROIs. For each ROI, the mean INT values are shown by group on the bottom of the boxplot, and the significant pairwise p values from post hoc comparisons are given on the top (threshold at 0.05). Age, sex, MFD, and CT were adjusted. The Tukey–Cramer method was applied to adjust for comparing a family of four estimates. Data corresponding to the comparison results are included in Extended Data Figure 5-1.

Figure 5-1

Extended Data for Figure 5. Pairwise comparison results of the cortical ROIs in which the INT values in MCI group are significantly shorter than other groups. ES: effect size, which is measured by Cohen's d statistics; df: degree of freedom. The unadjusted p value, Tukey-Cramer adjusted p value and the q value by FDR correction are provided. Download Figure 5-1, XLSX file ^{(11.8KB, xlsx)} .

INT differences in subcortical regions

We investigated INT differences in eight subcortical regions, including the hippocampus, a structure known to be compromised functionally and structurally in AD. As shown in Figure 6, four ROIs had significant group differences, which are the caudate, hippocampus, pallium, and putamen. In general, the AD and Converter groups tended to have longer INT values than the CN and MCI groups. The contrasts are particularly apparent between the MCI and Converter groups (MCI–Converter, caudate, $\mathrm{estimate},-0.058$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.533$; $t_{(897)}=-3.643$; $p=0.0016$; hippocampus, $\mathrm{estimate},-0.042$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.428$; $t_{(897)}=-2.924$; $p=0.0185$; pallidum, $\mathrm{estimate},-0.041$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.461$; $t_{(897)}=-3.147$; $p=0.0092$). The full comparison results can be found in Extended Data Figure 6-1.

Figure 6.

Subcortical ROIs in which the INT values have significant pairwise group differences. A, Brain visualization of the ROIs. The colormap indicate the significance of group differences in INT by ANOVA test (p values and q values from FDR correction are provided). B, The boxplots of the INT values by diagnosis groups on the identified subcortical ROIs. For each ROI, the mean INT values are shown by group on the bottom of the boxplot, and the p value of the significant pairwise difference is given on the top. Age, sex, and MFD were adjusted. The Tukey–Cramer method was applied to adjust for comparing a family of four estimates. Data corresponding to the comparison results are included in Extended Data Figure 6-1.

Figure 6-1

Extended Data for Figure 6. Pairwise comparison results of the subcortical ROIs in which the INT values in AD/Converter groups are significantly longer than stable MCI group. ES: effect size, which is measured by Cohen's d statistics; df: degree of freedom. The unadjusted p value and Tukey-Cramer adjusted p value are provided. Download Figure 6-1, XLSX file ^{(11.6KB, xlsx)} .

Converter group shows similar INT profile as AD group

In Figure 7, the INT contrast is minimal between AD and Converter groups but mostly distinct between MCI and MCI groups. As shown in Figures 4–6, no significant differences have been found between AD and Converter groups. For the five ROIs that showed significantly longer INTs in the Converter group than the MCI group, four of them also had significantly longer INTs in the AD group (except for the hippocampus). This includes fusiform (MCI–Converter, $\mathrm{estimate},-0.068$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.458$; $t_{(896)}=-3.084$; $p=0.011$), inferior temporal (MCI–Converter, $\mathrm{estimate},-0.081$; {\rm Cohen's $d=-0.562$; $t_{(896)}=-3.784$; $p=0.0009$), middle temporal (MCI–Converter, $\mathrm{estimate},-0.077$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.425$; $t_{(896)}=-2.868$; $p=0.0219$), and superior temporal (MCI–Converter, $\mathrm{estimate},-0.075$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.421$; $t_{(896)}=-2.852$; $p=0.023$) areas. Three ROIs have longer INT values in the Converter group than in the CN group, which are inferior temporal (CN–Converter, $\mathrm{estimate},-0.075$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.517$; $t_{(896)}=-3.590$; $p=0.002$), caudate (CN–Converter, $\mathrm{estimate},-0.040$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.366$; $t_{(897)}=-2.596$; $p=0.047$), and pallidum areas (CN–Converter, $\mathrm{estimate},-0.035$; $\mathrm{Cohe}{\mathrm{n}}^{\mathrm{\prime }}\mathrm{s}d=-0.391$; $t_{(897)}=-2.778$; $p=0.028$).

Figure 7.

Contrast INT maps between Converter and CN, MCI, and AD groups, respectively. The effects of age, sex, MFD, and CT were adjusted.

The pairwise-comparison results and the effect sizes (mean INT values) indicate that the INT alterations in these areas in the Converter group are already very similar to the AD group at the baseline, suggesting similar pathological changes prior to clinical presentation of AD.

Discussion

We used rs-fMRI data to investigate INTs of individuals in CN, MCI, and AD populations including those who convert to AD. INT is closely related to brain functional hierarchy (Murray et al., 2014; Hasson et al., 2015). Using established cortical INTs from an independent young and healthy adult sample as a functional index of hierarchy, we found that (1) AD and Converter groups are similar, as they both had had longer INTs in low-hierarchical-order areas compared with CN, and (2) stable MCI is distinct from AD and Converter groups, which had a less pronounced hierarchical gradient effect compared with CN with shorter INTs in high-hierarchical-order areas.

The pairwise-comparison results showed that distinct cortical INT alterations in AD were present in the entorhinal, fusiform, and inferior temporal areas; subcortical alterations were in the caudate, hippocampus, pallidum, and putamen. Three ROIs had significant INT alterations in the Converter group: the inferior temporal, caudate, and pallidum areas. Specific differences between MCI and AD/Converter groups were present in the fusiform, inferior temporal, middle temporal, precentral, and superior temporal areas.

Our results suggest distinct pathophysiological changes in the stable MCI and AD/Converter groups, at least as they relate to INT. For the AD and Converter groups, neural information is stored for a longer time in lower-hierarchical-order areas, while higher levels of the hierarchy seems to be preferentially impaired in stable MCI. Interestingly and perhaps counterintuitively, INT in primary somatosensory and motor cortices (BA 1, 2, 3, and 4, precentral and postcentral ROIs) and primary visual cortex (BA 17, pericalcarine) were lowest in the stable MCI group, perhaps an indication of non-AD aging changes in this group. The similarity in INT between the AD and Converter groups, and the differences between the MCI and Converter groups, suggests the potential for INT as a biomarker to predict conversion from MCI to AD. Future studies with larger samples of Converters are needed to develop INT—potentially in conjunction with other neuroimaging measures and/or cognitive and clinical panels—as a biomarker of conversion using multivariate machine learning methods.

Prior biophysical modeling work has indicated that INT depends on the strength of recurrent excitation within a brain region. Therefore, our observation of increased INT in the AD and Converter groups suggests increased E/I. Converging evidence suggests that E/I imbalance is a critical regulator of AD pathology (Palop et al., 2007; Giovannetti and Fuhrmann, 2019; Bi et al., 2020). Studies have shown that GABAergic dysfunction in AD impinges on the function of inhibitory neurons and on their ability to orchestrate and balance excitatory neurons (Giovannetti and Fuhrmann, 2019), and an imbalance of excitation and inhibition will lead to epileptogenesis and AD pathogenesis (Palop et al., 2007; Vossel et al., 2017)—both of which point to increased E/I in AD. The similarity of INT values between the Converter and AD groups indicates that pathological E/I imbalance already started before the conversion to AD. Thus, identifying regions with disrupted E/I balance areas using INT in AD and Converter groups constitutes an important first step to develop new diagnostic techniques and new treatment paradigms that aim to restore E/I imbalance for early intervention.

Altered INT was observed in several cortical areas implicated in the early stages of AD pathology. The entorhinal cortex is a medial temporal lobe brain region that often exhibits the earliest histological alterations in AD, including the formation of neurofibrillary tangles and cell death accompanied with deficits of memory and spatial navigation (Igarashi, 2022). The inferior temporal gyrus plays an important role in the mediation of verbal fluency; such cognitive function is affected early in the onset of AD (Scheff et al., 2011). Among subcortical regions, the hippocampus is the most extensively studied brain area in AD, with rapid neurodegeneration in the early stage of AD, which may be associated with the functional disconnection to other parts of the brain (Kantarci, 2005; Rao et al., 2022). Interestingly, in healthy individuals, the hippocampus has a low hierarchical order, indicating that its INT is short. This finding also demonstrates that despite substantial structural atrophy, INT is not shortened but rather lengthened compared with controls. Abnormalities in INT in the basal ganglia were also identified. While being underappreciated, studies have found atrophy of the caudate nucleus and other parts of the basal ganglia in the preclinical and clinical stages of AD (Miklossy, 2011; Elshafey et al., 2014; Persson et al., 2018). Consistent with current literature, these areas with altered INT may represent a stable biomarker for clinical diagnosis and an important therapeutic target in AD.

What we found was both different than our hypothesis and more complex yet neurobiologically plausible. In the MCI and Converter and AD groups, there were increases in duration in areas with known short INTs and decreases in areas that are known to have longer INTs as found by interaction examination of the intercept and slope. Additionally, several temporal lobe regions had higher INTs in the Converter and AD groups than the stable MCI and CN groups in planned regional contrasts. This result was in keeping with the aforementioned interaction effects. We view this pattern as a compensatory response to ongoing neurodegenerative changes that include synaptic loss and cell death.

One strength of the study is that it is the largest rs-fMRI sample from ADNI in the literature. With a sample size of 945, it is well powered to detect INT alterations in terms of both hierarchical gradient and individual regions. Furthermore, recent work has suggested that INT may have greater precision (i.e., test–retest reliability) as a measurement than other resting-state network metrics, an important point in favor of its use, especially to the extent that it makes the work more replicable (Shinn et al., 2023). With respect to limitations, the categorization of the diagnosis groups is based on the baseline and end-visit clinical status. The overall follow-up period is either up to 10 years or until participants left the study early. While follow-up duration does not significantly differ across groups, subjects in the stable MCI groups may potentially convert to AD after leaving the study, especially if the follow-up interval is too short; yet, such outcomes remain uncertain. Moreover, the sample is largely Caucasian and may or may not generalize to other ethnic groups. Additionally, our study was intentionally limited to investigating the effects of global severity (CN and AD) and progression/stable MCI subtypes. Future work will focus on the following aspects: (1) examining INT in the normal aging population and comparing INT differences with those in CN individuals; (2) understanding the relationships between INT and other AD biomarkers, including Aβ and phosphorylated tau and its effect on cognitive test performance; and (3) exploring the relationship between INT and the severity of symptoms or behavioral measurements in AD, such as the Mini-Mental State Examination and Functional Activities Questionnaire.