Spatially and temporally progressive hypoperfusion in Alzheimer's disease revealed by normative modeling

Abstract

INTRODUCTION

Cerebral perfusion is implicated in Alzheimer's disease (AD), but its development in AD and mild cognitive impairment (MCI) is not well characterized.

METHODS

We constructed a normative model using > 12,000 arterial spin labeling MRI scans and applied generalized additive models for location, scale, and shape (GAMLSS). Individual deviation z scores were derived by normative model, and outlier regions (z ≤ 2.3) were quantified as the total negative proportion (TNP) of extreme hypoperfusion. These metrics were then related to other AD biomarkers through linear modeling.

RESULTS

Compared to cognitively normal controls, AD showed higher TNP and greater longitudinal increases (p = 0.003), indicating progressive hypoperfusion. Progressive MCI exhibited greater perfusion decline than stable MCI (p = 0.01). Perfusion changes correlated with cognition, brain volume, amyloid, and apolipoprotein E status (all p < 0.05).

DISCUSSION

Normative modeling revealed inter‐individual heterogeneity in cerebral perfusion trajectories, underscoring its potential relevance for AD development.

Highlights

• This is the first normative modeling of longitudinal cerebral perfusion in Alzheimer's disease (AD).

• Our large‐scale multi‐dataset modeling provides strong evidence of variability.

• Hypoperfusion propagation emerges as a stable marker of AD progression.

• Perfusion‐based markers are associated with conversion from mild cognitive impairment to AD.

• Perfusion changes correlate with cognition, amyloid, volume, and apolipoprotein E genotype.

1. BACKGROUND

Alzheimer's disease (AD) is a progressive neurodegenerative disorder that affects millions worldwide and is the most common cause of dementia. ¹ , ² , ³ A prevailing hypothesis suggests that AD originates from amyloid beta (Aβ) accumulation, which triggers tau pathology, leading to neurodegeneration and cognitive impairment. ¹ , ³ , ⁴ , ⁵ This amyloid cascade model has guided much of the AD research, driving advances in biomarker discovery and therapeutic development.

However, cognitively normal individuals may exhibit amyloid levels comparable to those in preclinical AD, and reducing amyloid burden has not consistently improved cognition. ⁶ These findings suggest that mechanisms beyond amyloid plague depositions may contribute to AD. Among these factors, cerebral blood flow (CBF), the delivery of oxygen and nutrients essential for normal brain function and health, has gained increasing attention. ⁷ Emerging evidence indicates that age‐related changes in CBF occur during normal aging and become more pronounced in AD. ⁸ , ⁹ , ¹⁰ CBF alterations in aging and AD may result from vascular stiffening, reduced vessel density, cardiovascular dysfunction, and Aβ‐related vascular pathology, all of which can contribute to a chronic reduction in cerebral blood supply. ¹¹ , ¹² Given the brain's high dependence on efficient blood flow, disrupted CBF may lead to early neurovascular dysfunction, impaired clearance of toxic proteins, and progressive cognitive decline. ⁹ , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ While most studies report age‐related CBF decline, some have observed transient regional increases during early cognitive decline before subsequent hypoperfusion, underscoring the non‐linear and heterogeneous nature of perfusion changes in aging. Understanding normative CBF patterns across the lifespan and their deviations in AD is essential for elucidating disease mechanisms.

Normative modeling provides a statistical framework to quantify individual deviations relative to population expectations and to account for cohort‐level heterogeneity. ²⁰ , ²¹ It has been successfully applied to neurodevelopmental, ²² psychiatric, ²¹ , ²³ neurodegenerative disorders, ²⁴ , ²⁵ serving as a valuable tool for identifying disorder‐specific deviations at the individual level. Beyond cross‐sectional applications, incorporating longitudinal data enables tracking individual trajectories and disease progression. ²⁶ , ²⁷

Previous normative modeling studies in AD have primarily focused on structural brain changes. ²⁴ , ²⁵ , ²⁸ However, brain atrophy tends to be subtle in early AD and mild cognitive impairment (MCI), as structural changes typically emerge later in the disease course, posing challenges for early diagnosis. ¹ In contrast, CBF may be more sensitive to early AD‐related changes, ¹⁰ , ²⁹ given its essential role in maintaining brain metabolism and vascular health. It also plays a critical role in AD pathogenesis and serves as a potential regional functional biomarker. Moreover, reduced CBF is consistently linked to cognitive decline and disease severity. ⁶ , ³⁰ , ³¹ In addition, regional heterogeneity in perfusion decline may parallel volumetric atrophy patterns, which diverge between typical and atypical AD subtypes. ³² Given its sensitivity to early functional changes in AD, integrating cerebral perfusion into a normative modeling framework may better capture individual variability in AD and MCI.

CBF can be non‐invasively measured with arterial spin labeling (ASL) perfusion magnetic resonance imaging (MRI). ³³ , ³⁴ Over the past decades, ASL MRI has been increasingly used in AD research and reliably detected the AD‐related hypoperfusion patterns in the temporoparietal cortex. ¹⁸ , ¹⁹ , ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ Longitudinal CBF changes in aging and AD have been reported in two studies, ¹⁸ , ¹⁹ but the normative longitudinal CBF change patterns have yet to be systematically examined.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using PubMed and relevant Alzheimer's disease (AD) neuroimaging studies. While a few studies have investigated longitudinal cerebral perfusion changes in AD and mild cognitive impairment (MCI), no prior work has examined individual‐level perfusion trajectories using normative modeling. Relevant citations are appropriately referenced.

2. Interpretation: This study is the first to apply large‐scale normative modeling to longitudinal cerebral perfusion data in AD and MCI. Findings reveal substantial individual variability and suggest that hypoperfusion propagation is a stable marker of AD progression. Normative modeling offers a sensitive, individualized approach beyond traditional group‐level analyses.

3. Future directions: Future research should examine: (a) differences in cerebral perfusion progression between stable controls and those converting to MCI, to capture the earliest stages of cognitive decline; and (b) the clinical utility of normative modeling tools for individualized risk assessment.

In this study, a normative model using longitudinal ASL perfusion MRI data was constructed to quantify individual cerebral perfusion trajectories in AD and MCI. The model uses generalized additive models for location, scale, and shape (GAMLSS), a flexible statistical framework endorsed by the World Health Organization for modeling non‐linear growth trajectories and harmonizing batch effects across multi‐site data. ⁴² , ⁴³ , ⁴⁴ By incorporating multi‐site ASL perfusion data from > 10,000 scans as a reference dataset, along with disease dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI), this normative model provides a robust and sensitive tool for detecting atypical cerebral perfusion changes in AD and MCI. Specifically, we aim to: (1) track disease trajectories in individuals with MCI and AD; (2) investigate cerebral perfusion differences among MCI subtypes; and (3) examine associations between cerebral perfusion changes and other biomarkers, including cognition, brain structure, amyloid burden, and apolipoprotein E (APOE) genotype.

2. METHODS

2.1. Participants and dataset

To investigate cerebral perfusion changes in AD, this study used quality‐controlled structural MRI and ASL data. Only participants with both ASL and T1‐weighted images available and passing quality control were included. Data quality was assessed by visual inspection and quantitative quality evaluation index (QEI) values, and scans with QEI ≤ 0.1 or visible artifacts were excluded. ⁴⁵ Participants were drawn from two datasets: a reference (training) dataset and a disease dataset. The reference dataset consisted of healthy individuals across multiple cross‐sectional and longitudinal cohorts, including data from the Prevent AD and Dallas Lifespan studies, serving as a normative baseline. ⁴⁶ In total, the reference dataset comprised 12,430 scans (see Table S1 in Supporting Information for details) covering ages 0.75 to 88.7 years, with 53.8% female and 46.2% male. The disease dataset was obtained from ADNI and included 213 cognitively normal (CN; 649 scans), 256 individuals with amnestic MCI (911 scans: note that ADNI specifically excluded vascular MCI), and 64 patients with AD (189 scans). Participants in this dataset ranged from 54 to 89 years, with 50.4% female and 49.6% male. The diagnosis label of participants was obtained from the ADNI database, based on established clinical criteria. All ASL sequence information of the included datasets is provided in Table S2 in Supporting Information for details.

2.2. Image preprocessing

2.2.1. Structural data preprocessing

Structural MRI images were processed using the publicly available HCP pipeline. ⁴⁷ The preprocessed T1‐weighted images were used as structural references for CBF estimation and image registration, and were also used to extract brain volume.

2.2.2. ASL data preprocessing

The ASL MRI data were processed using ASLtbx ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ following standard ASL preprocessing pipelines. The ASL label and control images were motion corrected and spatially smoothed with a 6 mm full width at half‐maximum (FWHM) Gaussian kernel before CBF quantification. Quantitative CBF maps were calculated in physiological units (mL/100 g/minute) using the general kinetic model:

$$CBF=\frac{\mathrm{\Delta }M}{2\ast M0\ast \lambda \ast \alpha \ast \beta \ast e^{-\sigma /T}},$$ (1)

where $\mathrm{\Delta }M$ is the difference between control and label images, $M0$ is the equilibrium magnetization of arterial blood, $\lambda$ is the brain–blood partition coefficient, $\alpha$ is the labeling efficiency, $\beta$ is the bolus duration, $\sigma$ is the post‐labeling delay (PLD), and $T$ ​is the longitudinal relaxation time of blood (1.65 seconds at 3T). For 3D background‐suppressed ASL and pulsed ASL (PASL) data, M₀ images were explicitly acquired. In contrast, for 2D background‐unsuppressed pseudo‐continuous ASL (pCASL) data, the mean of the control images from the ASL time series was used as the M₀ image. Partial volume correction was performed using tissue segmentation results derived from the structural T1‐weighted images, following the same method. ³⁹ Quantitative CBF maps were calculated using the single‐compartment model recommended by the ASL consensus guidelines (Alsop et al.³⁶), assuming complete delivery of labeled blood water to the tissue compartment. For multi‐PLD acquisitions, CBF was estimated by fitting the ASL signals to the general kinetic model. ⁴⁸ , ⁴⁹ , ⁵⁰

Functional‐to‐standard image registration was performed to align processed CBF maps to the MNI152 standard brain space for region‐specific analyses. First, the CBF map was linearly registered to the T1‐weighted structural image, followed by non‐linear registration of the T1‐weighted image to the MNI152 brain template. The final transformation applied the combined functional‐to‐structural affine matrix and non‐linear warp. All registration outputs underwent visual inspection to ensure accuracy. All registration outputs were visually inspected to ensure accuracy. Regional CBF values were measured using the Harvard–Oxford brain atlas thresholded at 50% probability, encompassing 48 cortical and 8 subcortical regions. ⁵² Additionally, network‐level CBF values were explored using Yeo's atlas. ⁵³ It is noted that the field‐of‐view limitations of 2D ASL may lead to underestimation of CBF in the visual network when analyzed using Yeo's atlas.

2.3. Normative model

2.3.1. Model construction

To estimate individual cerebral perfusion deviation scores, GAMLSS using the gamlss package (version 5.0‐6) in R 4.2.0 was applied. ⁴² , ⁴³ The procedure involved two steps: selecting the optimal data distribution and determining the best‐fitting model parameters. ⁵⁴ Global cortical CBF was used to select the optimal distributions among all continuous distribution families available in the gamlss package. The Johnson's Su (JSU) distribution was identified as the optimal one based on the lowest Bayesian information criterion. ⁴⁶

The GAMLSS framework was applied with CBF as the dependent variable, age as a smoothing term (using B‐spline basis functions), and sex as a fixed effect. ASL sequence type was included as a fixed effect, while dataset site was treated as a random effect. ⁵⁴ The JSU distribution, which has four parameters: median (μ), coefficient of variation (σ), skewness (ν), and kurtosis (τ), was chosen to fit the data distribution. Each CBF value, denoted by y, was modeled as:

$$\gamma =JSU\left(\mu ,\sigma ,\nu ,\tau \right)$$ (2)

$$\mu =f_{\mu }\left(age\right)+{\beta }_{\mu }^1\left(sex\right)+{\beta }_{\mu }^2\left(sequence\right)+Z_{\mu }\left(site\right)$$ (3)

$$\sigma =f_{\sigma }\left(age\right)+{\beta }_{\sigma }\left(sex\right)$$ (4)

$$\nu ={\beta }_{\nu }$$ (5)

$$\tau ={\beta }_{\tau }$$ (6)

Following prior studies, only intercept terms were included for the ν and τ parameters. ²⁶ For model estimation, default convergence criterion of log‐likelihood < 0.001 between iterations was used, with a maximum of 200 iteration cycles. The global cortical, white matter, subcortical regions, along with network level and regional level CBF, were used to construct the normative model. The normative cerebral perfusion chart and model stability results from split‐half validation with 1000 repetitions are presented in Figures S1 and S2, respectively, in Supporting Information.

2.3.2. Individual deviation score

A key advantage of normative modeling is its emphasis on incorporating site‐matched controls in the testing set to mitigate site effects. To ensure robust modeling, a stratified approach was used, integrating site‐specific CN individuals in the testing set to minimize confounding with site effects in case–control comparisons. CNs from disease‐specific datasets were randomly split into training and testing subsets. The lifespan normative model was constructed using the training set, which included CNs and data from reference datasets (Table S1). The testing set, consisting of the remaining half CNs and all patient cases, served as an independent validation set for calculating deviation scores. Quantile scores relative to the normative model curves were calculated for each individual in each model, followed by the computation of deviation z scores. These z scores were derived by transforming the fitted quantiles into standard Gaussian z scores using quantile randomized residuals. ⁵⁵ The testing set (remaining CNs and AD, MCI) served as an independent validation set for calculating deviation z scores. These were computed by transforming quantiles into standard Gaussian z scores via quantile randomized residuals, averaged across 100 iterations. The standard deviation of individual deviation scores was calculated across 100 iterations to assess the stability of the models. In this AD/MCI cohort, deviations were predominantly negative, so extreme negative deviations were defined as z ≤ 2.3. ⁵⁴ Percentage maps of extreme deviations highlighted substantial heterogeneity in AD and MCI. For each subject, the total negative proportion (TNP, expressed as a percentage) was calculated as the proportion of extreme negative deviations (z ≤ 2.3) across 56 brain regions (48 cortical and 8 subcortical regions). ²⁵

2.4. MCI subtype definition

MCI is a highly heterogeneous group, necessitating an exploration of cerebral perfusion changes across subtypes. According to the diagnosis label, stable MCI included individuals who remained MCI throughout follow‐up (average maximum follow‐up: 37.2 months), while progressive MCI referred to those who converted to AD (maximum follow‐up: 35.5 months; average time to conversion: 26.0 months). This classification allowed for the investigation of cerebral perfusion changes associated with disease progression from MCI to AD.

2.5. Cerebral perfusion progression model

Cerebral perfusion progression within each group was explored. The time point of each scan was recoded as the number of months since the first scan and categorized into yearly intervals (± 3 months). To assess long‐term changes, cerebral perfusion at baseline and 2 years or later was compared. Linear mixed models (LMMs) were used to analyze the effect of time on cerebral perfusion metrics. The dependent variables included TNP and z scores for global, network, and regional perfusion, with time (in months) as the independent variable. Group differences were assessed by modeling the interaction between group and time effects. Age, sex, and baseline perfusion level were included as covariates, and individual participants were treated as random effects to account for repeated measures over time.

2.6. The associations of cerebral perfusion with other biomarkers

Based on prior research and the vascular hypothesis of AD, changes in cerebral perfusion during normal aging and AD progression are associated with disease severity and AD‐related pathophysiological processes. The beta values for each participant were calculated to quantify longitudinal changes in cerebral perfusion metrics (TNP and individual deviation scores of whole cortical regions), Clinical Dementia Rating Sum of Boxes (CDR‐SB) scores, and gray matter volume (GMV). Given the predictive role of amyloid deposition and the limited availability of longitudinal amyloid data (florbetapir standardized uptake value ratio [SUVR]), we examined the influence of baseline amyloid burden on subsequent perfusion changes. Preprocessed AV45 positron emission tomography (PET) SUVR data from the UC Berkeley ADNI pipeline were used for amyloid quantification. To assess the relationships between changes in cerebral perfusion and other bio‐marker alterations, linear models adjusting for age and sex as covariates were applied. ²⁵

APOE ε4 status was classified into three groups: homozygous carriers (ε4/ε4), heterozygous carriers (ε4/ε3), and non‐carriers (ε3/ε3). Group differences in the rate of change in TNP and cerebral perfusion z score were assessed using linear regression, with age and sex included as covariates in all models.

To evaluate the behavioral relevance of network‐level perfusion alterations, we examined associations between network‐based deviation scores and domain‐specific cognitive performance derived from the ADNI dataset. For each participant, cognitive indices were computed across six domains: episodic memory (mean of delayed recall, total learning, and forgetting ratio in Rey Auditory Verbal Learning Test [RAVLT]), executive function (mean of Trail Making Test Parts A/B, time‐reversed), language (Boston Naming Test), attention (Mini‐Mental State Examination [MMSE] attention subscore), visuospatial construction (Alzheimer's Disease Assessment Scale‐Cognitive subscale Constructional Praxis), and global cognition (MMSE total score).

2.7. MCI to AD progression analysis

Survival analysis with the Cox proportional hazard models was used to examine the association between baseline TNP and the risk of progression from MCI to AD. Given that the baseline TNP distribution in MCI was highly skewed with most values equal to zero, participants were dichotomized into TNP = 0 (low TNP) and TNP > 0 (high TNP) groups, consistent with the approach used by previous work. ²⁵ A Kaplan–Meier plot was used to visualize the progression from MCI to AD. Meanwhile, a logistic regression analysis, with a binomial distribution, was conducted to assess the association between baseline TNP subtype and the risk of conversion from MCI to AD during the follow‐up period. Age and sex were included as covariates to account for potential confounding factors.

2.8. Classification and prediction in MCI

Support vector machine (SVM) was used to classify stable MCI and progressive MCI based on baseline cerebral perfusion deviation scores. Additionally, support vector regression (SVR) was used to predict the time to AD conversion in progressive MCI cases. More details on the classification and prediction models can be found in our previous study. ⁴⁶ All models were repeated 1000 times using a 2‐fold cross‐validation framework, alternating training and testing sets in each fold. Area under the curve (AUC) significance was assessed via a non‐parametric permutation test (1000 iterations) with randomized labels before classification or prediction.

2.9. Sensitivity analysis

To evaluate the robustness of our findings to the choice of outlier threshold, sensitivity analysis was performed to use a less stringent cutoff of z ≤ 1.96 as extreme negative deviation to calculate the TNP instead of the primary threshold of z ≤ 2.3. ²⁵

To examine whether amyloid pathology and APOE ε4 genotype influenced the longitudinal perfusion trajectories, ⁵⁶ additional sensitivity analyses were conducted by incorporating amyloid PET SUVR and APOE ε4 carrier status as additional covariates within the longitudinal mixed‐effects models. We further tested whether amyloid burden or APOE genotype modified the rate of perfusion change by including their interactions with time (Time x Amyloid, Time x APOE) in the model. Participant‐specific random intercepts were modeled to account for repeated measures, and age and sex were retained as covariates.

2.10. Statistical analysis

All statistical analyses were performed with R 4.2.0. Group differences in age were assessed using analysis of variance, while sex and APOE‐negative proportion were analyzed using a chi‐squared test. Longitudinal progression was examined using LMMs, with time as a within‐subject factor and group as a between‐subject factor. Age and sex were included as covariates to control for confounding effects. False discovery rate (FDR) correction was applied separately for global cerebral perfusion (four measurements × six group comparisons) and network (six measurements × six group comparisons) cerebral perfusion.

3. RESULTS

3.1. The characteristics of participants

Two hundred thirteen CN, 251 MCI, and 64 AD patients were included, with a total of 649, 911, and 129 scans, respectively (Table 1). The maximum follow‐up time was comparable between CN and MCI but shorter in AD. Age and sex distributions differed across groups. Additionally, the three groups exhibited significant differences in cognition (CDR‐SB and MMSE), amyloid PET SUVR, GMV, and APOE ε4 carrier status.

TABLE 1.

Demographics and clinical data in disease dataset.

	CN	MCI	AD	Statistical
n	213	251	64
Total scan	649	911	129
Max follow‐up time (months)	29 ± 27.2	29.2 ± 25.1	11.8 ± 7.92
Sex (female: male)	127:86	111:140	28:36	X 2 = 14.68, p = 0.0021
Baseline age (years)	73.24 ± 7.48	71.29 ± 7.33	74.85 ± 7.24	F = 8.776, p = 0.0002
Baseline MMSE score	29.0 ± 1.20	27.9 ± 2.17	22.6 ± 2.78	F = 171.4, p < 0.0001
MMSE change rate (score/year)	β = 0.0001, p = 0.596	β = –0.035, p < 0.0001	β = –0.23, p < 0.0001	F = 52.219, p < 0.0001
Baseline CDR‐SB score	0.135 ± 0.628	1.57 ± 1.28	4.63 ± 2.00	F = 224, p < 0.0001
CDR‐SB change rate (score/year)	β = 0.0013, p = 0.209	β = 0.02, p < 0.0001	β = 0.114, p < 0.0001	F = 33.54, p < 0.0001
Baseline amyloid PET SUVR	0.739 ± 0.119	0.839 ± 0.170	0.97 ± 0.166	F = 35.25, p < 0.0001
Baseline GMV (mm3)	583,390 ± 52,633	593,606 ± 52,461	562,404 ± 57,730	F = 6.761, p < 0.0001
GMV change rate (mm3/year)	β = –24.95, p = 0.61	β =  –115.77, p = 0.0175	β = –836.21, p < 0.0001	F = 15.47, p < 0.0001
APOE ε4 non‐carrier (percentage in group)	72.5%	52.1%	31.8%	X2 = 32.06, p < 0.0001

Abbreviations: AD, Alzheimer's disease; APOE, apolipoprotein E; CDR‐SB, Clinical Dementia Rating Sum of Boxes score; CN, cognitively normal; GMV, gray matter volume; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; PET, positron emission tomography; SUVR, standardized uptake value ratio.

3.2. Longitudinal cerebral perfusion in AD

Figure 1 shows the distinct longitudinal cerebral perfusion trajectories of AD, MCI, and CN, highlighting significant differences in their progression patterns (Table S3 in Supporting Information).

FIGURE 1.

Longitudinal perfusion progression in AD. A, Regional longitudinal perfusion deviation scores: maps depict perfusion deviation scores across cortical regions in CN, MCI, and AD groups at baseline and follow‐up. B, Violin plots of perfusion changes: perfusion changes across CN, MCI, and AD groups displayed globally and within specific brain networks. C, Cerebral perfusion changes and group differences: comparisons of TNP percentage changes and cerebral perfusion changes among CN, MCI, and AD groups. D, T‐value maps illustrating longitudinal cerebral perfusion changes within each group. For TNP, the total negative proportion was calculated as the proportion of extreme negative deviations (z < –2.3) across 56 brain regions. AD, Alzheimer's disease; AN, affective network; CN, cognitively normal; DAN, dorsal attention network; DMN, default mode network; FPN, frontoparietal network; MCI, mild cognitive impairment; SSN, somatomotor network; SUB, sub‐cortical regions; TNP, total negative proportion; VAN, ventral attention network; VN, visual network; WM, white matter.

At baseline, the TNP of cerebral perfusion deviations (Z ≤ 2.3) was significantly higher in AD compared to CN (estimate = 3.33, t = 2.56, p = 0.011, p FDR = 0.038), while no significant difference was observed in MCI (p = 0.067). Over time, TNP increased significantly in AD relative to CN (estimate = 3.223, t = 3.148, p = 0.00169, p FDR = 0.01), whereas MCI did not show a significant increase (p = 0.454). Comparisons between AD and MCI showed no baseline differences in TNP (p = 0.17), but AD exhibited a significantly faster increase in TNP over time (estimate = 3.208, t = 2.964, p = 0.0031, p FDR = 0.015), suggesting a more rapid spatial spread of perfusion abnormalities in AD (Figure 1C).

At the global level, including cortical, white matter, and subcortical regions, AD showed significantly greater baseline deviations compared to CN in both cortical (estimate = –0.79, t = 4.78, p < 0.000001, p FDR < 0.00001) and subcortical regions (estimate = –0.64, t = 4.052, p = 0.00005, p FDR = 0.0006). Similarly, significant baseline differences were observed between AD and MCI in cortical (estimate = –0.501, t = 3.17, p = 0.0017, p FDR = 0.01) and subcortical regions (estimate = –0.42, t = 2.78, p = 0.0057, p FDR = 0.027). However, no significant group differences were found in the longitudinal rate of perfusion decline across AD, MCI, and CN (all p > 0.1; Figure 1C).

At the network level, AD exhibited significantly reduced baseline perfusion in the visual, dorsal attention, frontoparietal, and default mode networks compared to CN and MCI (p < 0.05, FDR corrected; Table S3, Figure 1B). In contrast, no significant group differences were observed in the longitudinal rate of perfusion decline across networks (all p > 0.4). Additionally, AD exhibited marked baseline perfusion differences (compared to CN and MCI) and distinct progression patterns over time in regional levels (Figures 1A, 1D and S3 and S4 in Supporting Information).

3.3. Longitudinal cerebral perfusion in MCI subtype

The demographic and clinical information of two types of MCI subtype (diagnostic outcome, Figure 2A) are presented in Table S4 in Supporting Information.

FIGURE 2.

Longitudinal perfusion progression in stable MCI and progressive MCI. A, Stable MCI was defined as remaining in MCI status while the MCI individuals who convert to AD later were defined as Progressive MCI. B, Regional longitudinal perfusion deviation scores: maps depict perfusion deviation scores across cortical regions in stable MCI and progressive at baseline and follow‐up. C, Violin plots of perfusion changes: perfusion changes displayed globally and within specific brain networks. D, Cerebral perfusion changes and group differences: comparisons of TNP percentage changes and cerebral perfusion changes among stable MCI and progressive MCI. E, T‐value maps illustrating longitudinal cerebral perfusion changes within each group. MCI_S: stable MCI; MCI_P: progressive MCI. AD, Alzheimer's disease; AN, affective network; DMN, default mode network; FPN, frontoparietal network; MCI, mild cognitive impairment; SSN, somatomotor network; SUB, sub‐cortical regions; TNP, total negative proportion; VAN, ventral attention network; VN, visual network; WM, white matter.

In Figure 2D, at baseline, no significant difference in TNP was observed between progressive MCI and stable MCI (p = 0.367). However, progressive MCI demonstrated a significantly faster increase in TNP over time compared to stable MCI (estimate = 1.593, t = 3.913, p = 0.001, p FDR = 0.01). At the global cerebral perfusion level, no baseline and progression differences in cortex, white matter, subcortical regions were observed between the two groups. No significant differences were observed in baseline or longitudinal changes of network‐level cerebral perfusion between progressive and stable MCI (p > 0.1). The global and network cerebral perfusion changes are presented in Figure 2C,D and Table S5 in Supporting Information. Lastly, MCI subtypes exhibited distinct progression patterns over time in regional levels (Figures 2B, 2E).

3.4. Associations between cerebral perfusion changes and cognitive, amyloid, and brain volume markers

Figure 3 illustrates the associations between cerebral perfusion change rates and other key biomarkers. Across all participants (Figure 3A), changes in cognition, as measured by the rate of change in CDR‐SB scores, were significantly associated with changes in both TNP (t = 3.197, p = 0.0015) and cortical cerebral perfusion (t = –2.400, p = 0.0169). Baseline amyloid deposition (Figure 3B) was also significantly associated with subsequent changes in TNP (t = 2.816, p = 0.0054) and cortical perfusion (t = –2.359, p = 0.0193). Additionally, changes in brain volume (Figure 3C) were significantly correlated with TNP changes (t = –2.497, p = 0.0132), although no significant association was found with cortical CBF (p = 0.64).

FIGURE 3.

Associations between cerebral perfusion change rates and key biomarkers. A, Relationships between cerebral perfusion change rates and cognitive decline, indexed by changes in CDR‐SB scores. B, Associations between baseline amyloid burden (Amyloid SUVR) and subsequent cerebral perfusion changes. C, Associations between changes in brain volume and cerebral perfusion change rates. CDR‐SB, Clinical Dementia Rating Sum of Boxes, SUVR, standardized uptake value ratio.

3.5. Associations between cerebral perfusion and APOE ε4 status

Figure 4 shows cerebral perfusion changes over time across different APOE genotypes in global, network and regional levels (Figure 4A, B). Compared to non‐carriers, both APOE ε4 homozygous (estimate = 0.146, t = 2.965, p = 0.0031) individuals and APOE ε4 heterozygous (estimate = 0.06, t = 2.655, p = 0.008) exhibited significantly higher TNP. No significant differences in TNP were observed among CN individuals based on APOE status, whereas a substantial difference was found in MCI with APOE ε4 homozygous (estimate = 0.185, t = 2.934, p = 0.0035) participants.

FIGURE 4.

Longitudinal cerebral perfusion progression by APOE status. A, Violin plots of perfusion changes: perfusion changes displayed globally and within specific brain networks. B, T‐value maps depicting longitudinal cerebral perfusion differences between APOE status groups. C, Comparison of TNP changes and cerebral perfusion differences by APOE genotype. AN, xxxx; APOE, apolipoprotein E; DMN, default mode network; FPN, frontoparietal network; MCI, mild cognitive impairment; SSN, somatomotor network; SUB, sub‐cortical regions; TNP, total negative proportion; VAN, ventral attention network; VN, visual network; WM, white matter.

Regarding the global cerebral perfusion change rate (Figure 4C), no significant differences were detected between APOE ε4 homozygous individuals and non‐carriers across the cortex (p = 0.412), white matter (p = 0.464), or subcortical regions (p = 0.367).

After FDR correction, significant positive associations were observed between baseline deviation score in the dorsal attention, frontoparietal control, default mode, and visual networks and multiple cognitive domains, particularly episodic memory, executive function, and global cognition (Figure S5 in Supporting Information). These relationships indicate that preserved perfusion within these large‐scale networks supports better cognitive performance in individuals across the AD spectrum.

3.6. Clinical implication of cerebral perfusion deviation in MCI

Survival analysis showed that each regional perfusion outlier (increase 2 unit in TNP) in baseline perfusion was associated with a 4% increased risk of progression from MCI to AD (hazard ratio = 1.022, 95% confidence interval [CI]: [1.003, 1.042], p = 0.02; Figure 5A). The logistic regression analysis suggested a trend toward an increased risk of AD conversion in the high baseline TNP group compared to the low baseline group (p = 0.09). We evaluated classification performance between stable MCI (n = 133, age: 71.4 ± 7.11) and progressive MCI (n = 59, age: 72.8 ± 7.51) using baseline individual deviation scores with 1000 permutation tests, achieving 71.3% accuracy (95% CI: [69.5%, 73.2%]), a mean AUC of 0.589 (95% CI: [0.581, 0.595]), and statistical significance (p = 0.02; Figure 5B). Additionally, perfusion deviations significantly predicted the time to conversion from progressive MCI to AD (R ² = 0.148, 95% CI: [0.0275, 0.2016], p = 0.0031; Figure 5C).

FIGURE 5.

Clinical application of cerebral perfusion in MCI prognosis. A, Kaplan–Meier survival plot comparing the proportion of remaining MCI patients between the low and high baseline TNP MCI groups. B, Classification performance distinguishing stable MCI from progressive MCI. C, Predictive value of perfusion deviation scores for the time‐to‐conversion from progressive MCI to AD. AD, Alzheimer's disease; MCI, mild cognitive impairment; TNP, total negative proportion.

3.7. Sensitivity analysis

To assess the sensitivity of our results to threshold of defining TNP, we repeated the analyses using a less stringent cutoff (z ≤ 1.96, Figure S6 in Supporting Information). At baseline, both MCI (p = 0.035) and AD (p = 0.0005) showed significantly higher TNP of hypoperfusion compared to CN. Longitudinally, only AD exhibited a significant TNP increase relative to CN (p = 0.02), whereas MCI showed no significant progression effect (p = 0.146).

Incorporating amyloid PET SUVR and APOE ε4 carrier status as additional covariates did not alter the main findings. The group x time interaction remained significant after adjusting for these variables (time x AD vs. time x CN: p = 0.0029; time x AD vs. time x MCI: p = 0.0079). No significant time x amyloid and time x APOE interaction effects were found. The group x time interaction between AD and CN remained significant (p = 0.02) after including the time x amyloid burden and time x APOE interactions in the same model. These results indicate that the longitudinal CBF trajectories were not solely driven by amyloid burden or genetic risk.

Across 100 independent repetitions, individual deviation estimates remained highly consistent, with the mean standard deviation across repetitions < 0.09, demonstrating excellent split‐half stability of deviation z scores under model re‐training (Figure S7 in Supporting Information).

4. DISCUSSION

We presented the first study to construct the normative model to explore the cerebral perfusion progression in AD using the largest ASL MRI dataset to date (> 10,000 scans). Individualized CBF change trajectories during MCI and AD were captured, and the individualized cerebral perfusion deviation scores and outlier maps (TNP) illustrated the non‐uniform impact of AD on patients as the disease progresses. Moreover, our findings reveal notable heterogeneity within MCI, indicating that cerebral perfusion decline accompanies disease progression toward AD. Furthermore, cerebral perfusion changes were associated with key biomarkers, including brain volume, cognition, amyloid deposition, and APOE genotype, suggesting a multi‐omics mutual interplay in AD pathophysiology.

Our findings provide novel insights into the progression of cerebral perfusion abnormalities in AD, emphasizing a regionally inhomogeneous and propagative pattern of perfusion decline. Consistent with previous studies, ³¹ , ⁵⁷ we observed significant baseline perfusion deficits in AD compared to CN. However, unlike the expected accelerated global CBF reduction, our longitudinal analysis, aligned with a previous study, ⁵⁸ revealed no significant difference in the overall rate of perfusion decline between groups. However, AD patients showed a greater increase in the number of brain regions with severely reduced CBF, as reflected by the progressive rise in TNP over time compared to CN and MCI. This rapid spatial spread of hypoperfusion may be a key driver of neurodegeneration‐related cognitive decline. ¹⁰ , ¹² It suggests that the expansion of outlier regions, rather than the overall rate of cerebral perfusion decline, may better explain progressive cognitive deterioration in AD. ¹⁰ , ⁵⁹ , ⁶⁰ Certainly, this should be further validated with ASL data acquired with a state‐of‐the‐art ASL sequence.

Notably, previous studies have shown that AD pathology is characterized by the transneuronal spread of tau pathology along functional and structural brain networks, leading to neuronal damage, inflammation, and progressive cognitive impairment. ⁶¹ , ⁶² , ⁶³ , ⁶⁴ In parallel, several studies have reported a close association between cerebral perfusion and tau pathology in AD. ⁶⁵ , ⁶⁶ The rapid expansion of spatial hypoperfusion patterns may impair tau clearance across widespread regions, contributing to progressive cognitive decline across multiple domains.

Our findings indicate that progressive MCI exhibits greater cerebral perfusion changes than stable MCI, suggesting a potential role for CBF dynamics in the progression toward AD. Previous studies have identified key regions, including the precuneus and posterior cingulate cortex, where cerebral perfusion alterations are associated with the progression from MCI to AD. ⁴⁰ , ⁵⁸ , ⁶⁷ , ⁶⁸ In line with these findings, our regional maps (Figures 2E and S3 and S4) revealed hypoperfusion in posterior cortical areas overlapping the precuneus and posterior cingulate cortex in progressive MCI. Our results further highlight the potential of cerebral perfusion as a biomarker, as baseline TNP levels were associated with a higher risk of conversion to AD in survival analysis. Additionally, cerebral perfusion measures exhibited a modest but statistically significant ability to differentiate stable from progressive MCI (AUC = 0.59, p = 0.02), indicating that baseline perfusion alterations provide informative, though limited, discriminative value. These findings indicate that cerebral perfusion alterations may reflect early vascular vulnerability along the AD continuum, providing complementary information for risk stratification and disease monitoring.

Our findings highlight the intricate interplay between cerebral perfusion changes and key aging‐related biomarkers, including cognitive decline, amyloid deposition, brain atrophy, and APOE genotype. Emerging evidence suggests that cerebral perfusion may play an early role in AD progression by compromising the brain's energy supply and impairing the clearance of metabolic waste. ⁶⁸ Reduced CBF has been linked to blood–brain barrier (BBB) dysfunction, which diminishes clearance efficiency and promotes the accumulation of pathological proteins. ⁶⁸ , ⁶⁹ Over time, these vascular impairments can lead to synaptic dysfunction, neuronal injury, and white matter damage, ultimately contributing to brain atrophy and cognitive decline. In our study, higher baseline amyloid deposition was associated with greater longitudinal declines in cerebral perfusion, consistent with previous findings that amyloid accumulation may contribute to neurovascular dysfunction. ¹ , ³ , ⁴ These results support the presence of a complex, bidirectional relationship between vascular and amyloid pathologies in AD. APOE ε4 is known to impact cerebrovascular function, contributing to BBB and blood–CSF barrier dysfunction, increased capillary permeability, and cerebral amyloid angiopathy. ⁷⁰ , ⁷¹ Our findings indicate that APOE ε4 carriers show greater perfusion abnormalities, and that the ε4 allele's long‐term vascular effects may contribute to the progressive spread of perfusion disruptions.

Compared to previous studies that primarily focused on group differences in AD, our study is the first to apply a normative modeling framework, offering several key advantages and novel insights. First, normative modeling enables individualized inference, capturing disease heterogeneity beyond traditional group‐based comparisons. By leveraging large multisite datasets, this approach provides more precise estimates of individual deviations from normative trajectories. Importantly, it not only quantifies the distribution of deviation scores to reflect heterogeneity but also allows for identifying the proportion of extreme deviation maps. Second, by explicitly modeling biological and site‐related variability, the normative model reduces unexplained variance and increases sensitivity to subtle perfusion deviations. Notably, 13.8% of MCI patients exhibited at least three brain regions with outlier perfusion values, which was nearly twice of the proportion observed in CN (6.4%) at baseline, underscoring the potential utility of perfusion deviation in early disease detection.

Several challenges and limitations should be acknowledged in this study. First, the 2D PASL data in the disease cohort were acquired with a suboptimal ASL sequence with relatively lower signal‐to‐noise ratio at a single post‐labeling delay, limiting more fine‐grained exploration of vascular and regional perfusion characteristics in AD. Moreover, susceptibility and slice‐coverage artifacts, particularly in posterior visual and inferior temporal regions, may still influence the accuracy of regional CBF quantification. Although sequence type and acquisition parameters were accounted for in the modeling framework, residual differences in ASL acquisition (e.g., spatial resolution/voxel size) cannot be fully accounted for and may contribute to additional variability. Second, the biological and physiological variables (e.g., hematocrit, blood pressure) may influence CBF quantification, ⁷² but these measures were not consistently available across cohorts and could not be modeled; this remains a limitation common to large‐scale ASL studies. Advanced denoising through traditional machine learning or deep learning may help mitigate these confounds but those methods still need more data to establish generalizability. ⁷³ , ⁷⁴ , ⁷⁵ , ⁷⁶ , ⁷⁷ , ⁷⁸ Third, while ADNI excluded major cerebrovascular disease, mild vascular risk factors such as hypertension or diabetes may still be present in older adults. Moreover, cardiovascular risk information was not consistently available across all training datasets, precluding explicit adjustment for vascular burden. Future studies integrating cardiovascular and metabolic health measures could help refine vascular‐adjusted normative models. Fourth, the limited availability of longitudinal ASL–PET data precluded testing the causal direction between hypoperfusion and amyloid/tau pathology. Existing evidence suggests a bidirectional relationship, ¹⁰ , ¹² , ⁶⁸ and future multimodal longitudinal studies are needed to disentangle whether vascular dysfunction precedes or follows protein aggregation. Fifth, similar to prior normative modeling studies using structural MRI, ⁴⁴ our neuroimaging datasets (both reference and disease cohorts) predominantly represent populations from Europe and North America. Although the reference dataset was primarily composed of healthy volunteers, it is difficult to completely exclude individuals with subclinical or comorbid conditions (e.g., vascular risk, early neurodegeneration), especially in large population‐based cohorts such as UK Biobank. This residual heterogeneity should be considered a limitation of the normative model. To improve the generalizability of perfusion‐based biomarkers in AD, future studies should prioritize the inclusion of more diverse neuroimaging cohorts, ensuring broader ethnic representation and reducing geographic bias.

5. CONCLUSIONS

This study used a normative modeling framework to characterize longitudinal cerebral perfusion changes in AD and MCI at the individual level, revealing interindividual variability and disease progression patterns. Perfusion abnormalities propagate over time and are closely linked to key biomarkers and genotype, highlighting its potential as an early biomarker for AD conversion risk.

CONFLICT OF INTEREST STATEMENT

All authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All participants provided written informed consent in accordance with the institutional review board policies of the contributing studies and the ADNI consortium.

Supporting information

Supporting Information

Supporting Information

Zeng X, Li Y, Hua L, et al. Spatially and temporally progressive hypoperfusion in Alzheimer's disease revealed by normative modeling. Alzheimer's Dement. 2026;22:e71203. 10.1002/alz.71203