Longitudinal changes in subcortical functional connectivity during Alzheimer’s disease progression

Abstract

Human cognition and behavior rely on the integration of large-scale neural networks that connect the cerebral cortex and subcortical structures. Emerging evidence suggests that alterations in the functional connectivity (FC) between the cortical and subcortical regions in Alzheimer’s disease (AD) may influence the onset and progression of both cognitive and noncognitive symptoms at the group level. However, an individualized and longitudinal framework to capture deviations in subcortico-cortical FC from normative brain aging remains underexplored. We addressed this gap by leveraging large-scale longitudinal neuroimaging datasets and applying a normative modeling approach to characterize subcortical FC trajectories across the adult lifespan. First, we quantified individual deviations in the subcortical FC in individuals with cognitive impairment (CI) relative to a normative aging group using centile scores and tracked longitudinal changes across multiple follow-ups. We examined the relationship between changes in subcortical FC and clinical measures of cognitive function, including episodic memory, executive function, and language. Our findings revealed widespread decreases in the subcortical FC in individuals with CI, except in the limbic network, which diverged from the patterns observed in normal aging. These alterations are significantly associated with a decline in memory and executive functions. Collectively, our results may advance our understanding of AD-related connectopathy and provide a direction for profiling individualized longitudinal FC changes in individuals with CI. Furthermore, our results could inform individualized prognosis and targeted interventions.

1. Introduction

Human cognition and behavior arise from the synergistic operation of large-scale, intricately interconnected neural networks rather than the isolated functioning of discrete brain regions [7,28]. A growing body of evidence has demonstrated robust anatomical and functional connections between subcortical structures and various cortical regions, forming complex neural systems that support cognitive and behavioral processes [11,31]. For example, the formation of explicit memories depends on the integrity of the entorhinal cortex-hippocampal system [47]. Sensory information is relayed to the cortex via subcortical structures such as the thalamus, which plays a central role in integrating sensory inputs, coordinating motor outputs, regulating emotional states, and supporting higher-order cognitive functions [9,30]. The subcortical and cortical regions interact through multiple “top-down” and “bottom-up” pathways that modulate memory formation, arousal, emotional regulation, and attentional processes, underscoring the critical role of subcortico-cortical connectivity.

Alzheimer’s disease (AD) is a neurodegenerative disorder that progressively affects cognition, emotions, and quality of life [26,50]. Early stages of AD typically manifest as episodic memory impairment, followed by a decline in language, visuospatial abilities, and executive function [41]. To assess the brain changes related to these symptoms, studies have traditionally focused on the hippocampus, medial temporal lobes, and neocortical association areas because of their early involvement and established links to memory deficits [3,43]. However, this cortico-centric perspective may neglect the important role of subcortical structures, such as the striatum and thalamus, in AD pathophysiology. Given the extensive modulatory influence of the subcortical structures on cortical circuits, their dysfunction can disrupt information flow in brain networks [10,46], emphasizing the need to investigate subcortico-cortical vulnerability in AD.

Recent advances in neuroimaging have made it possible to examine the functional integrity of the distributed brain networks in vivo. Functional connectivity (FC), defined as the temporal correlation of spontaneous brain activity measured by resting-state functional magnetic resonance imaging (rs-fMRI), is increasingly recognized as a key feature in brain pathology [13,39]. FC offers a promising avenue for investigating AD pathology in clinical settings. Recent studies have shown that tau pathology spreads along functional networks from local epicenters, suggesting that connectivity mapping may help identify disease progression pathways [12,20]. Moreover, a recent FC study demonstrated that tau accumulation follows the sensory-association cortical hierarchy and that altered functional modular organization is linked to cognitive decline, providing quantifiable fMRI-based markers for AD [34]. Other studies have used FC measures to distinguish normal aging from AD-specific brain alterations [38] and to predict conversion from mild cognitive impairment (MCI) to AD [24]. Furthermore, when FC is combined with other biomarkers, such as amyloid positron emission tomography (PET) or microtubule-associated protein tau (MAPT) gene expression, it explains substantially more variance in tau accumulation than any single modality alone [35] and effectively identifies risk factors for rapid disease progression [33]. Together, FC-based analyses hold significant potential for improving our understanding of AD pathology. Thus, tracking longitudinal FC changes can provide insights into the progression of neurodegeneration and its cognitive impact.

Most existing studies have focused on group-level comparisons, which limit their ability to capture individual and longitudinal deviations from normative aging trajectories. This limitation makes distinguishing between pathological changes and normal variability challenging, requiring individualized analytical approaches to better quantify network disruptions in AD. Normative modeling mitigates these limitations by establishing population-level reference curves of age-related brain changes and mapping individual deviations from these trajectories. This approach enables personalized detection of atypical subcortical connectivity patterns while preserving the heterogeneity inherent to disease progression [4,15,27,29,49].

The present study investigated subcortical FC alterations in patients with AD using a normative modeling framework. Our primary objectives were to (1) map normative aging trajectories of the subcortical FC across the adult lifespan using large-scale healthy control datasets, (2) quantify individual deviations from these normative trajectories in individuals with AD, and (3) examine the relationship between subcortical FC changes and clinical measures of cognitive function, including episodic memory, executive functioning, and language abilities. We hypothesized that the subcortical FC in individuals with AD may exhibit atypical trajectories compared with normative aging and that these alterations may be associated with changes in cognitive function.

2. Methods

2.1. Study participants

We collected and analyzed the MRI data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database of 146 individuals with cognitively normal (CN) and 142 individuals with cognitive impairment (CI), including MCI and AD. Individuals with CI underwent multiple scans (between three and 13), resulting in 656 longitudinal data points. Individuals with CI were included based on the following criteria: (1) availability of both T1-weighted (T1w) and rs-fMRI scans, (2) without diagnostic reversion (e.g., changes from AD to MCI), and (3) provision of at least three rs-fMRI sessions. Individuals with CN were included based on the following criteria: (1) availability of both T1w and rs-fMRI scans, (2) without diagnostic conversion (e.g., changes from CN to MCI) (3) provision of at least two rs-fMRI sessions. Finally, 563 longitudinal scans were obtained. The normative models were constructed using baseline data, and the follow-up scans were used for validation. Institutional Review Board (IRB) approval was obtained from the original study depicting ADNI. In the ADNI dataset, consent forms had been approved by each participating institution’s IRB. All ADNI data were fully anonymized, and no protected health information was included. In addition to the ADNI dataset, we used an independent cohort of 725 neurologically healthy individuals from the Human Connectome Project-Aging (HCP-A) database [6] to construct normative aging trajectories of the subcortical FC. The IRB of Washington University previously approved the participant recruitment procedures and informed consent forms, including the consent to share de-identified data. The demographic information of the study participants is summarized in Table 1.

Table 1.

Demographic information of the study participants.

	HCP-A	ADNI-CN	ADNI-CI	ADNI-CN vs. CI (p-value)
Number of subjects/scans	725	146/563	142/656	-
Baseline age (range)	60.4 ± 15.7 (36 - 100)	72.5 ± 6.41 (57.4 - 90.5)	73.9 ± 7.77 (55.9 - 91.5)	0.104a
Sex (M:F)	319:406	61:85	84:58	0.003b
Baseline MMSE	-	29.2 ± 1.01	27.9 ± 1.75	<0.001a
APOE4 carrier (non-carrier:carrier)	-	96:50	79:63	0.079b
Education	-	16.6 ± 2.27	16.2 ± 2.67	0.132a
Number of scans during follow-ups (range)	-	4.17 ± 1.78 (2 - 9)	3.38 ± 2.04 (3 - 13)	0.001a

Mean and standard deviation are reported if applicable.

Abbreviations: HCP-A, Human Connectome Project-Aging; ADNI, Alzheimer’s Disease Neuroimaging Initiative; CN, cognitively normal; CI, cognitive impairment; M, male; F, female; MMSE, Mini-Mental State Examination; APOE4, apolipoprotein E4.

^aTwo-sample t-test.

^bChi-squared test.

2.2. MRI data acquisition

• i) ADNI: We obtained the T1w MRI data from the ADNI database, scanned using 3T scanners. T1w images were scanned with a three-dimensional (3D) magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence (Philips: repetition time [TR] = 6.8 ms, echo time [TE] = 3.16 ms, flip angle = 9°, voxel size = 1$\times$1$\times$1.2 mm), an accelerated sagittal MPRAGE (Siemens: TR = 2300 ms, TE = 2.95 ms, flip angle = 9°, voxel size = 1 mm isotropic), and an accelerated sagittal IR-FSPGR (GE: TR = 7.36 ms, TE = 3.06 ms, flip angle = 11°, voxel size = 1 mm isotropic). Fluid-attenuated inversion recovery (FLAIR) images were acquired from the ADNI-GO/2 phase, using a two-dimensional (2D) FLAIR sequence (TR = 9000 ms, TE = 90 ms, inversion time [TI] = 2500 ms, flip angle = 150°). Rs-fMRI data were obtained using a 2D echo planar imaging (EPI) sequence (TR = 3000 ms, TE = 30 ms, flip angle = 80°, 140 volumes, voxel size = 3.31 mm isotropic). FLAIR images were acquired from the ADNI-3 phase using a 3D FLAIR sequence (TR = 4800 ms, TE = 119 ms, TI = 1650 ms, flip angle = 120°), and rs-fMRI with an EPI sequence (TR = 3000 ms, TE = 30 ms, flip angle = 90°, 197 volumes, voxel size = 3.4 mm isotropic).

• ii) HCP-A: T1w and T2w MRI were acquired using a Siemens Prisma 3T scanner with MPRAGE and T2-SPACE sequences, respectively (T1w: TR = 2500 ms; TE = 1.81, 3.60, 5.39, or 7.18 ms; FOV = 256 × 256 mm; voxel size = 0.8 mm isotropic; T2w: TR = 3200 ms; TE = 564 ms; FOV = 256 × 256 mm; voxel size = 0.8 mm isotropic). Rs-fMRI data were obtained using a Siemens Prisma 3T scanner with a 2D multiband gradient-recalled echo EPI sequence (TR = 800 ms, TE = 37 ms, flip angle = 52°; 478 volumes; voxel size = 2 mm, isotropic). During the two sessions, two runs were performed in the opposite phase-encoding directions (anterior-posterior [AP] and posterior-anterior [PA]).

2.3. MRI data preprocessing

• i) ADNI: The MRI data were preprocessed using fMRIPrep version 23.2.0 [16]. Intensity non-uniformity corrections were applied to the T1w images, followed by skull stripping. A T1w reference map was generated by registering multiple T1w images scanned across longitudinal sessions. Brain tissue segmentation, including cerebrospinal fluid (CSF), white matter, and gray matter, was performed using FSL [23]. Brain surface reconstruction was performed using FreeSurfer [18], with a FLAIR image employed to refine the pial surface delineation. Volume-based spatial normalization to the Montreal Neurological Institute (MNI) standard space was performed via nonlinear registration using Advanced Normalization Tools (ANTs) [1]. Preprocessing was performed for each rs-fMRI session. A reference volume was generated from the middle frame, and head motion corrections were performed. After slice-timing correction, the reference volume was registered to the T1w reference map using six degrees of freedom and subsequently registered to the standard space. Cortical surface models based on pial and white matter boundaries, along with subcortical segmentation masks, were used to generate a grayordinate space [21]. Preprocessed fMRI data were projected onto a standard grayordinate space using a cortical ribbon-constrained volume-to-surface mapping algorithm. The first 10-second volumes were censored to allow for magnetic field saturation, and volumes with severe head motion (i.e., framewise displacement [FD] > 0.5 mm) were discarded. A bandpass filter (0.008–0.1 Hz) was applied, and the effects of the six head motion parameters, white matter, and CSF were regressed out.

• ii) HCP-A: The imaging data were preprocessed using the HCP minimal preprocessing pipeline [21]. The T1w and T2w data underwent gradient nonlinearity and b0 distortion corrections, and the images were coregistered using a rigid-body transformation. Bias field correction was performed based on the inverse intensities from T1- and T2-weighting. The processed data were registered nonlinearly in the MNI standard space. For rs-fMRI data, preprocessing included distortion correction, head motion correction, bias field correction, intensity normalization, volume-to-surface mapping to standard grayordinate space using multimodal surface matching (MSMAll), high-pass filtering (>0.009 Hz), ICA-FIX-based denoising, and spatial smoothing with a 2 mm full width at half maximum (FWHM) Gaussian kernel.

2.4. PET data acquisition and preprocessing

We acquired PET images from the ADNI-GO/2/3 databases, which were processed using standardized dynamic protocols. The amyloid-β PET consisted of AV45- and FBB-PET. For AV45-PET, a protocol lasting 50–70 min post-intravenous injection of 370 MBq of [18F] Florbetapir was used, with a scan duration of 20 min divided into four 5 min frames. For the FBB-PET, the protocol lasted 90–110 min post-intravenous injection of 300 MBq of [18F] Florbetaben, with a 20-min scan divided into four 5 min frames. For the tau-PET, a 75–105 min protocol post-injection of 370 MBq of [18F] AV-1451 (Flortaucipir) was applied, with a 30 min scan divided into six 5 min frames. All PET data were matched to corresponding fMRI sessions. The PET images were preprocessed as follows. Raw PET images were coregistered to correct for head motion, and the aligned frames were averaged over each 5-min interval. The resulting images were reoriented to conform to a standard image grid of 160 × 160 × 96 matrix with a voxel size of 1.5 mm³ voxel size, followed by an intensity normalization. Spatial smoothing was applied using an 8 mm FWHM. The PET images were registered to the corresponding T1w image using a rigid-body transformation and subsequently nonlinearly registered onto the standard MNI space. The standardized uptake value ratio (SUVR) for subcortical regions was calculated using the cerebellum as the reference region. For the amyloid PET, we transformed their SUVR into centiloid (CL) units using the established transformation formula (https://adni.loni.usc.edu/wp-content/themes/freshnews-dev-v2/documents/pet/ADNI%20Centiloids%20Final.pdf) to facilitate the comparability of AV45-PET and FBB-PET measurements [25].

2.5. Seed-based FC

The subcortical structures of the accumbens, amygdala, caudate, hippocampus, pallidum, putamen, and thalamus were defined using FreeSurfer, and the cortical regions were defined using the Schaefer atlas with 200 parcels [42]. We generated seed-based FC maps with the seeds from bilateral subcortical structures using preprocessed rs-fMRI data (Fig. 1A left). The subcortico-cortical and subcortico-subcortical FC were derived by calculating Pearson’s correlation between the mean time series of the seed and target regions (Fig. 1A right). The correlation coefficients were transformed using Fisher’s r-to-z to ensure normality of the data.

Fig. 1.

Schematic illustration of the study. (A) We constructed subcortico-cortical and subcortico-subcortical FC using rs-fMRI time series data. (B) Normative trajectory of subcortical FC was generated using individuals with CN, and the deviations of individuals with CI were quantified by calculating centile scores. The centile scores were calculated across longitudinal scans. (C) Longitudinal changes in subcortical FC centile scores were statistically assessed using linear mixed-effects models. For the regions showing significant effects, the associations between changes in subcortical FC centile scores and neuropsychological assessment scores were assessed.

Abbreviations: FC, functional connectivity; CN, cognitively normal; CI, cognitive impairment; ADNI, Alzheimer’s Disease Neuroimaging Initiative; ROI, region of interest.

2.6. Normative age trajectory and centile score estimation

A normative model of the subcortical FC was constructed using the HCP-A dataset. We opted for a generalized additive model for location, scale, and shape (GAMLSS), which allowed us to capture nonlinear age-related changes in the FC (Fig. 1B left) [48]. The GAMLSS model fits a four-parameter distribution model of µ (mean), σ (variance), ν (skewness), and τ (kurtosis) using the sinh-arcsinh (SHASH) distribution. Specifically, normative modeling of FC was performed for each subcortical region as a function of age and the interaction between age and sex as follows:

$$\text{FC}\sim \text{SHASH}\left(\mu ,\sigma ,\nu ,\tau \right),$$

$$\mu ={\beta }_{\mu }+{\beta }_{\mu ,\text{age}}f\left(\text{age}\right)+{\beta }_{\mu ,\text{age}*\text{sex}}f\left(\text{age}*\text{sex}\right),$$

$$\text{log}\left(\sigma \right)={\beta }_{\sigma }+{\beta }_{\sigma ,\text{age}}f\left(\text{age}\right),$$

$$\text{log}\left(\nu \right)={\beta }_{\nu },$$

$$\text{log}\left(\tau \right)={\beta }_{\tau },$$ (1)

where $f$ is a nonlinear function (i.e., P-spline) and $\beta$ denotes the regression coefficient. The GAMLSS model fitting was implemented using the gamlss R package [48]. The HCP-A dataset-derived normative model was transferred to the ADNI-CN group by re-estimating the first- and second-order intercepts (i.e., ${\beta }_{\mu }$, ${\beta }_{\sigma }$), while the nonlinear effect of age was fixed [14]. This procedure allowed the intercept and standard deviation (SD) of the model to be adapted to the new dataset while preserving the estimated nonlinear age-related effects. To quantify deviations in the subcortical FC of individuals with CI from the normative trajectory established in the ADNI-CN group, we calculated longitudinal centile scores (Fig. 1B right). The centile scores were determined by comparing each individual’s subcortical FC values to the normative distribution and identifying their relative positions within this distribution, providing age- and sex-specific measures of deviation across ages. Lower scores indicated hypoconnectivity, whereas higher scores reflected hyperconnectivity in individuals with CI relative to those in the CN group. To assess the reliability of the normative models across the subcortical regions, we evaluated the goodness-of-fit by calculating the absolute mean residuals for each subcortical connectivity.

2.7. Changes in subcortical FC during AD progression

For each subcortical region, we used a linear mixed-effects model to evaluate the time effects on the subcortical FC centile scores, controlling for MMSE, APOE4 status, and education level as follows:

$$\text{FC}\text{centile}\text{scor}{\mathrm{e}}_{ij}={\beta }_0+{\beta }_1\text{Tim}{\mathrm{e}}_{ij}+{\beta }_2\text{MMS}{\mathrm{E}}_{ij}+{\beta }_3\text{APOE}{4}_i+{\beta }_4\text{Ed}{\mathrm{u}}_i+z_i+{\epsilon }_{ij}$$ (2)

where $\beta$ values are the fixed effects, $z_i$ represents the random intercept for individual $i$, accounting for inter-individual variability, $j$ denotes the subcortical region, and ${\epsilon }_{ij}$ is the error term. In this analysis, age and sex were not included as covariates as their effects were already accounted for when computing the centile scores. We quantified the effect of time using the t-statistics of ${\beta }_1$ (i.e., the interval between baseline and follow-up), and stratified these effects across seven intrinsic functional networks, including visual, somatomotor, dorsal attention, ventral attention, limbic, frontoparietal, and default mode networks (Fig. 1C left) [51]. The multiple comparisons across the subcortical regions were corrected using the false discovery rate (FDR) procedure [2].

2.8. Associations between changes in subcortical FC and neuropsychological assessments

To assess the relationship between the longitudinal changes in the subcortical FC and cognitive performance, we conducted correlation analyses between annual changes in the subcortical FC centile scores and neuropsychological assessment scores for each participant. The annual change was estimated using linear regression applied to repeated measurements of both the FC centile scores and neuropsychological scores over time (Fig. 1C right). Neuropsychological assessments included composite scores for episodic memory (ADNI-MEM), executive function (ADNI-EF), and language (ADNI-LAN), which were centrally computed by the ADNI Neuropsychology Core Team according to standardized protocols (https://adni.bitbucket.io/reference/docs/UWNPSYCHSUM/adni_uwnpsychsum_doc_20200326.pdf). For each subcortical region, we calculated Pearson’s correlation between the changes in FC centile scores and those in each neuropsychological domain, focusing on regions identified as significant in previous analyses. This approach enabled us to quantify whether individuals with greater changes in subcortical FC experienced correspondingly slower or faster cognitive deterioration.

2.9. Associations between FC and PET biomarkers

To examine the relationship between longitudinal changes in subcortical connectivity and AD molecular pathology accumulation, we analyzed a subset of participants for whom longitudinal PET data (at least three timepoints) were available (n = 37 for amyloid, n = 22 for tau). For each subcortical region, we calculated the rate of longitudinal change in PET biomarkers, defined as the slope estimated from linear regression. For subcortical FC, we calculated degree centrality for each subcortical region from the connectivity matrix and computed longitudinal changes in the centile scores. We then computed the correlations between the changes in PET biomarkers (amyloid CL units and tau SUVR) and those in FC centile scores for each subcortical region. To account for the small sample size, the statistical significance was assessed using 200 permutation tests, where subjects were randomly shuffled to generate null distributions.

2.10. Sensitivity analyses

First, to evaluate the validity of transferring the HCP-A-based normative model to the ADNI dataset, we examined the residuals after model fitting. The goodness-of-fit was assessed using the absolute mean residuals for each subcortical connectivity, where smaller values indicate a better model fit. Furthermore, we constructed an additional normative model using only the CN cohort from ADNI. Centile scores of subcortical FC were calculated, and their longitudinal changes were assessed. The similarity between the HCP-A-transferred model and the ADNI-CN-only model was evaluated by computing the spatial correlation of the t-statistic maps. Second, to evaluate the validity of constructing a normative model using a cross-sectional dataset, we assessed whether the HCP-A-based model could predict longitudinal connectivity changes. Specifically, we used the HCP-A-based normative model to construct a prediction framework for estimating follow-up connectivity values in ADNI-CN participants. Prediction accuracy was quantified using the root mean squared error (RMSE) between the actual and predicted connectivity values. Additionally, we performed the same prediction analysis using a normative model constructed solely from ADNI-CN data to allow a fair comparison with the longitudinal ADNI-CN dataset. Third, to examine the effects of head motion on longitudinal changes in subcortical FC centile scores, we first tested the correlations between mean FD and subcortical FC values within the ADNI-CN and ADNI-CI groups, separately. We also evaluated the longitudinal effects of subcortical FC using linear mixed-effects models that included mean FD as a covariate. Fourth, we tested whether different sites influenced the longitudinal changes in subcortical FC centile scores by including site information as an additional covariate in the linear mixed-effects models. Lastly, to account for potential confounding effects due to the sex imbalance, we included sex as an additional covariate in the linear mixed-effects models.

3. Results

3.1. Normative age effects of subcortical FC

We calculated the subcortical FC for each individual in the HCP-A and ADNI cohorts (Fig. S1). When we assessed the spatial correlation between the mean FC values of the left and right seed regions in the HCP-A cohort, highly similar connectivity patterns were observed between the hemispheres, with a mean correlation of r = 0.982 ± 0.008 across all seeds, corroborating the rationale for using bilateral subcortical regions as seeds (Fig. S2A). We also found high correlations of subcortical FC between HCP-A and ADNI-CN (mean ± SD r = 0.723 ± 0.074), and between HCP-A and ADNI-CI (0.668 ± 0.102) (Fig. S2B). The difference between the ADNI-CN and ADNI-CI groups was statistically significant (p = 0.021). Next, we characterized age-related differences in the subcortical FC in seven bilateral subcortical regions using the HCP-A cohort (Fig. 1A), and distinct patterns were observed across the age groups (Fig. 2B). For example, the accumbens showed an overall increasing trend in FC with cortical networks with increasing age, whereas the hippocampal FC decreased. The amygdala and thalamus showed decreased FC with the somatomotor, visual, and attention networks. The caudate exhibited reduced FC with the frontoparietal and default mode networks, and the pallidum and putamen showed relatively weak increasing FC with the default mode network. These findings reflect the unique functional roles of each subcortical region. Comparable patterns of age-related changes were observed in the ADNI-CN cohort (Fig. S3). When we assessed the reliability of the normative model, small mean absolute residuals were observed for both HCP-A (mean ± SD = 0.005 ± 0.001 across subcortical structures) and ADNI datasets (0.024 ± 0.006), although the amygdala and pallidum exhibited slightly higher residuals (Fig. S4 and S5).

Fig. 2.

Normative aging snapshots of subcortical FC in the HCP-A cohort. (A) Brain maps show subcortical FC across the adult lifespan (ages 38–92). (B) Radar plots display network- and subcortical region-wise FC across ages.

Abbreviations: FC, functional connectivity; HCP-A, Human Connectome Project-Aging.

3.2. Longitudinal changes in subcortical FC centile scores in individuals with CI

To assess deviations in subcortical FC in individuals with CI from the normative reference trajectories (i.e., ADNI-CN), we calculated the centile scores for each individual at baseline and follow-up. Notably, widespread and progressive declines in centile scores were observed relative to normative expectations in all subcortical regions (Fig. S6). We quantitatively examined longitudinal changes in the FC centile scores during AD progression using linear mixed-effects models. We found that the caudate, hippocampus, putamen, and thalamus exhibited widespread and pronounced reductions in the centile scores over time, particularly in the visual, somatomotor, and attention networks (Fig. 3). In contrast, the orbitofrontal cortex and temporal pole, which are primarily involved in the limbic network, showed localized increases in the FC centile scores in the accumbens and caudate nucleus. Additionally, the medial occipital cortex exhibited increased FC centile scores in the hippocampus. These results indicate that CI is associated with longitudinal changes in the subcortical FC linked to widespread cortical networks.

Fig. 3.

Longitudinal effect of subcortical FC changes. Brain maps show longitudinal changes in subcortical FC centile scores for seven subcortical regions in the ADNI-CI cohort, with color indicating the t-statistic of the time effects. Bar plots summarize the t-statistics according to intrinsic functional networks.

Abbreviations: FC, functional connectivity; ADNI, Alzheimer’s Disease Neuroimaging Initiative; CI, cognitive impairment.

3.3. Associations between changes in the centile scores and neuropsychological assessments

We examined the clinical relevance of the subcortical FC alterations in individuals with CI by associating the changes in the centile scores and episodic memory performance for each brain region (Fig. 4). Overall, links of the somatosensory and dorsolateral prefrontal cortices with multiple subcortical regions showed significant positive correlations between centile scores and memory outcomes (r = 0.199 to 0.408, all p_FDR < 0.05), indicating that greater preservation (i.e., an increase) in subcortical FC over time may be associated with slower memory decline. However, significant negative correlations were observed in the precuneus, middle temporal cortex, and inferior parietal regions (r = −0.202 to −0.258, all p_FDR < 0.05), indicating that greater subcortical FC increase is linked to greater memory loss. Similarly, associations with executive function (ADNI-EF) and language (ADNI-LAN) scores revealed increases or decreases in subcortical FC, highlighting significant links between the subcortical FC changes and cognitive performance (Figs. S7-S8).

Fig. 4.

Associations between changes in subcortical FC centile scores and neuropsychological assessments. Brain maps show the correlations between annual changes in subcortical FC centile scores and changes in episodic memory performance (ADNI-MEM). Scatter plots illustrate the relationship between the changes in ADNI-MEM scores and mean changes in the centile scores that showed positive and negative associations, respectively.

Abbreviations: FC, functional connectivity; ADNI, Alzheimer’s Disease Neuroimaging Initiative.

3.4. Association between longitudinal changes in FC and PET biomarkers

We investigated whether the longitudinal progression of subcortical FC is associated with the accumulation of AD pathology by correlating longitudinal changes in subcortical FC and PET biomarkers. While limited statistical power prevented definitive conclusions, we observed negative associations between FC alterations and amyloid (mean r = −0.101 ± 0.157) and tau (mean r = −0.095 ± 0.102) accumulation. Specifically, a moderate trend was found in the accumbens (r = −0.317, p_perm = 0.055) and hippocampus for amyloid (r = −0.286, p_perm = 0.130; Fig. S9).

3.5. Sensitivity analyses

First, when we fitted the HCP-A-based normative model to the ADNI data, the residuals were small (mean ± SD = 0.024 ± 0.006 across subcortical structures) considering the range of their mean FC between −0.082 and 0.574, suggesting that the model transferred well despite differences in demographic and scanner characteristics between the datasets (Fig. S5). Moreover, when comparing the t-statistic maps of longitudinal changes in subcortical FC between the HCP-A-transferred model and the ADNI-CN-only model, we observed high spatial correlations across all subcortical regions (all r > 0.9; Fig. S10). Second, the HCP-A-based normative model successfully predicted longitudinal connectivity changes in ADNI-CN participants, with low mean RMSE values of 0.226 ± 0.023 across subcortical regions (Fig. S11), demonstrating that cross-sectional normative trajectories can effectively capture longitudinal variations in FC. When comparing the prediction results between the HCP-A-based model and the ADNI-CN-only model, the HCP-A-based model showed lower RMSE in most brain regions (i.e., green areas in Fig. S12). This demonstrates that large cross-sectional datasets could provide more general age trajectories. This likely reflects the advantage of leveraging the larger HCP-A dataset (n = 725) over the smaller ADNI-CN baseline dataset (n = 146) in capturing reliable population-level age trajectory patterns. Third, when we examined the relationships between mean FD and subcortical FC, no significant correlations were observed in any subcortical regions (all |r| < 0.13, all p_FDR > 0.7; Fig. S13). Moreover, the longitudinal effects derived from the FD-included model were highly consistent with those from the model without FD (all r > 0.9; Fig. S14), indicating that our findings were not driven by motion artifacts. Fourth, when we included site information as a covariate in linear mixed-effects models, the resulting t-statistic maps for longitudinal FC changes remained highly consistent with the original findings (all r > 0.9; Fig. S15). Lastly, when we included sex as a covariate, the t-statistic maps remained highly consistent with the original findings (all r > 0.9; Fig. S16).

4. Discussion

Subcortico-cortical circuits support essential cognitive and behavioral functions and may thus contribute to the progression of cognitive decline in neurodegenerative disorders. In this study, we used a normative modeling approach with longitudinal datasets to characterize age-related changes in the subcortical FC. We quantified the degree of deviation in the subcortical FC of each individual with CI from the normative aging trajectory and observed substantial abnormalities in multiple subcortical areas. As AD progressed, we observed widespread and progressive reductions in the FC of the hippocampus, thalamus, putamen, and caudate to the frontoparietal, visual, and attentional cortices. Additionally, localized increases in FC were identified linking the accumbens, caudate, and hippocampus to the limbic, visual, and default mode networks. The widespread FC decreases and localized increases tracked the longitudinal deterioration in episodic memory, executive function, and language, underscoring the clinical relevance of the degenerative and reorganizational processes of subcortical FC. These results suggest that AD pathology is associated with subcortico-cortical circuit disturbances.

Among the earliest and most consistent findings across the AD continuum is a widespread decline in the subcortico-cortical FC. Recent meta-analytic evidence highlights that subcortical regions show selective vulnerability in AD [45]. The thalamo-frontal circuit is of particular importance. A recent study reported that the FC between the mediodorsal thalamus and frontal lobes was significantly reduced in individuals with MCI and AD, whereas thalamic connectivity with occipital regions increased in those with AD [22]. These findings suggest that as the thalamo-frontal circuit breaks down, the thalamus may aberrantly shift its coupling toward the posterior part of the sensory network [22]. The striatum (caudate and putamen), another crucial neural component, forms loops with the frontal cortex and supports executive control, working memory, and motor planning [40]. Recent studies have shown that early dopaminergic dysfunction in the striatum progressively disrupts its connectivity with the limbic regions, suggesting a key role of frontostriatal circuits in AD pathophysiology [32,36,37]. Collectively, these findings highlight the importance of subcortico-cortical FC reduction in AD, and our results quantitatively support these prior observations.

In contrast to the widespread decrease in subcortical FC, we also observed localized increases in the FC centile scores in the visual and limbic networks. Notably, these increases involved connections between the orbitofrontal cortex and temporal pole with the caudate and accumbens and between the hippocampus and early visual cortex. The widespread FC decline, accompanied by increased FC, suggests divergent hyper- and hypoconnectivity patterns in subcortico-cortical circuits during AD progression. Hyperconnectivity in preclinical AD may reflect a compensatory mechanism that helps to preserve cognitive performance despite accumulating pathology [8]. For example, hyperconnectivity in the occipital regions has been reported in the prodromal phase of AD; this hyperconnectivity disappears within three years of dementia progression [5]. These findings suggest that local hyperconnectivity is transient and metabolically demanding. As the pathology progresses, this heightened activity becomes unsustainable, ultimately leading to large-scale hypoconnectivity and cognitive decline. Although further analysis is required to determine the causal relationship between FC changes and cognitive deterioration, our results underscore the importance of monitoring both hyper- and hypoconnectivity in the subcortico-cortical networks to better capture the trajectory of cognitive decline in AD.

However, the hyperconnectivity patterns need careful interpretation, as they may reflect either adaptive compensation or maladaptive reorganization. While hyperconnectivity is often conceptualized as an adaptive response that supports behavioral performance through compensatory recruitment of alternative networks, it may also represent maladaptive reorganization that emerges when compensatory capacity is exceeded [19]. In our analysis of associations with cognitive scores, we found that hyperconnectivity in certain regions was linked to worse cognitive performance, supporting the idea that disruptions of central hubs may amplify the brain’s response to pathology. In this context, our findings align with the biphasic trajectory model of AD, in which the interplay between amyloid and tau may give rise to both hypo- and hyperconnectivity patterns. It has been shown that individuals with amyloid positivity but low neocortical tau showed increased connectivity in the default mode and salience networks, whereas those with elevated tau exhibited decreased connectivity [44]. Furthermore, another study differentiated normal aging-related connectivity from pathology-related connectivity alterations, reporting that hypoconnectivity within the posteromedial cortex is associated with normal aging and cognitive decline, while medial temporal-posteromedial hyperconnectivity is specifically linked to early AD pathology, particularly in APOE4 carriers [17]. Similarly, we found that subcortico-cortical hyperconnectivity in limbic and visual regions occurred alongside progressive hypoconnectivity in other networks, suggesting distinct pathological processes rather than a uniform network response. The co-existence of hypo- and hyperconnectivity across different circuits underscores that AD involves both maladaptive reorganization and direct pathological disruption.

We investigated the association between changes in the subcortical FC and longitudinal cognitive outcomes. For episodic memory, greater preservation (or increase) of the subcortical FC over time in the somatosensory and dorsolateral prefrontal cortices was associated with slower memory decline, suggesting that maintaining the FC strength in these areas may support neuroprotective processes. Conversely, FC reduction in the middle temporal cortex, precuneus, and visual areas was associated with greater memory loss. Similar patterns were observed for executive and language functions. Notably, divergence in the subcortical FC does not necessarily indicate that increased FC is indicative of better cognitive symptoms. These findings indicate that properly characterizing the subcortico-cortical FC alterations in AD may help identify large-scale network disruptions that could guide the development of targeted interventions.

Our exploratory analysis of AD molecular pathology revealed negative associations between longitudinal changes in FC and amyloid accumulation in the accumbens and hippocampus. The negative correlations indicate that increased amyloid burden may contribute to progressive subcortico-cortical dysconnectivity. The lack of statistical significance may be due to the small subset of participants with sufficient longitudinal PET data. Future studies with larger samples of longitudinal PET data and longer follow-up periods are needed to establish clearer mechanistic links between subcortical connectivity alterations and molecular pathology progression in AD.

Despite its strengths, this study had several limitations that should be acknowledged. First, although the ADNI dataset is a valuable resource for studying brain-related AD pathology, the participants may not be fully representative of the broader AD population. Further validation using independent cohorts is required to enhance the generalizability of our results. Second, while our normative modeling approach successfully captured connectivity variations across most subcortical regions, certain structures, such as the amygdala and pallidum, exhibited slightly reduced model precision, which may be attributable to partial volume effects and low signal-to-noise ratios. Third, the CI group included patients in both MCI and AD stages. While this reflects the continuum of disease progression, the heterogeneity between the early (MCI) and late (dementia) stages may obscure stage-specific effects. Finally, unmeasured clinical factors (e.g., vascular risk factors) may influence both FC and cognitive outcomes; however, we could not account for these factors because of the limitations of the retrospective dataset.

In summary, our findings refine the understanding of AD-related connectopathy and outline the utility of longitudinally individualized FC trajectories for tracking disease progression. Together, our findings suggest that tracking longitudinal FC changes can provide valuable insights into the neurodegenerative process, thereby enhancing the potential for patient-specific prognosis and the development of targeted network-based interventions for CI. The results underscore the importance of actively integrating fMRI-based assessments into clinical practice.

Data availability

Imaging and phenotypic data were provided in part by the ADNI (https://adni.loni.usc.edu). ADNI was launched in 2003 as a public-private partnership led by principal investigator Michael W. Weiner. The primary goal of the ADNI is 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. The HCP-A datasets are publicly available from the Lifespan HCP Release 2.0 (https://www.humanconnectome.org/study/hcp-lifespan-aging).

Funding

This study was supported by the Institute for Information and Communications Technology Planning and Evaluation (IITP) funded by the Korea Government (MSIT) (No.2022–0–448/RS-2022-II220448, Deep Total Recall: Continual Learning for Human-Like Recall of Artificial Neural Networks; RS-2021-II212068, Artificial Intelligence Innovation Hub program), AI Graduate School Support Program (Sungkyunkwan University) (RS-2019-II190421), ICT Creative Consilience program (IITP-2025-RS-2020-II201821), and National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024–00408040, RS-2025–20253035).

Declaration of generative AI and AI-assisted technologies

The authors declare that no generative AI or AI-assisted technologies were used in the writing of this manuscript or in the creation of figures, images, or artwork.

CRediT authorship contribution statement

Sunghun Kim: Writing – original draft, Visualization, Validation, Methodology, Formal analysis, Data curation, Conceptualization. Sewook Oh: Writing – review & editing, Validation, Methodology. Hyunjin Park: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Bo-yong Park: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.