Divergent white matter metabolic signature patterns indicate impending cognitive decline in aging and dementia

Abstract

White matter (WM) is a key substrate for interregional neural communication and cognitive function but the role of WM glucose metabolism in cognitive aging has been understudied. Using multimodal neuroimaging (MRI, FDG-PET, amyloid-PET) from 3142 participants (15,287 visits) across two studies, we examined the contribution of WM to cognition and identified divergent WM signatures. Higher glucose metabolism in expected WM (EWM; corpus callosum and cingulum) was associated with better cognition, whereas increased metabolism in atypical WM (AWM; corona radiata) was linked to worse cognition, indicating a compensatory mechanism. EWM metabolism declined with aging, Alzheimer’s disease (AD) progression (amyloid-β and APOE-ε4 carrier), and white matter hyperintensities, while AWM metabolism increased with aging and vascular risk but was partially weakened by AD neuropathology. Longitudinally, higher EWM and lower AWM metabolism predicted slower cognitive decline. Divergent WM metabolic patterns shed light on the dynamic role of WM in maintaining cognitive function. This study emphasizes the complementary information provided by WM metabolism for predicting future cognitive decline and identifying cognitive resilience.

Divergent white matter metabolic patterns reveal a mechanistic basis for cognitive resilience and enhance prediction of future cognitive decline beyond established aging and dementia biomarkers.

Introduction

Ongoing research into the biological processes of Alzheimer’s disease (AD) has driven substantial advancements in its conceptual framework and the therapeutic development of treatments targeting core pathologies¹. However, the dissociation between AD neuropathology and clinical cognitive manifestations remains an unresolved challenge, as nearly 30% of older adults exhibit hallmark AD pathologies (amyloid-β (Aβ) plaques and neurofibrillary tangles) without concomitant cognitive impairment^2,3. This phenomenon, termed cognitive resilience, refers to the preservation of cognitive abilities despite age-related brain changes or AD-associated pathologies^4–6. In recent years, cognitive resilience has emerged as a critical focus in AD research. Investigating its underlying mechanisms may not only inform prevention strategies to delay cognitive decline but also refine the identification of high-risk populations for early intervention.

The brain mechanisms underlying cognitive resilience involve preserved brain structures, normal glucose metabolism, efficient network connectivity, and compensatory processes that buffer against age- and disease-related pathology^6,7. Gray matter (GM) supports local processing, synaptic plasticity, and high metabolic activity, with preserved cortical thickness, volume, and dendritic spine density in key regions (such as medial temporal and prefrontal cortices) linked to better cognition despite neurodegeneration^8–10. White matter (WM) ensures rapid communication between GM regions, with high microstructural integrity and connectivity enabling compensation for focal damage^8,11. Neuroglia (including astrocytes, microglia, and oligodendrocytes) maintain homeostasis, modulate injury responses, and preserve axonal health and myelination, thereby sustaining network efficiency and resilience^12,13. Metabolically, GM’s high glucose utilization supports synaptic activity, whereas WM’s lower but efficient energy metabolism maintains myelin and axons^14,15.

Fluorodeoxyglucose positron emission tomography (FDG-PET), which measures regional glucose metabolism and reflects neuronal/synaptic activity, astrocytic activity, and neuroinflammation¹⁶, is a valuable tool for investigating and quantifying brain regions crucial to cognitive resilience. Several studies have reported age- and dementia-related declines in brain glucose metabolism, particularly in certain GM regions, which are well-established features of AD and cerebrovascular disease^17,18. Our recent work suggests that FDG-PET uptake in the cingulate cortex and temporal pole could serve as a marker of cognitive resilience¹⁹. In contrast to these GM-centric findings, WM, despite its crucial role in cortical connectivity and cognitive integration^20,21, remains underexplored in metabolic studies. To date, relatively few studies have examined WM metabolism using FDG-PET, and most have focused on selected regions such as the corpus callosum, cingulum, or white matter hyperintensities (WMH) rather than the whole WM^15,22–29. A systematic characterization of WM metabolic patterns across the brain has not yet been established.

In this study, we used FDG-PET imaging with a specific focus on WM to characterize brain glucose metabolic patterns associated with aging, dementia, and cognitive performance. Our primary hypotheses were that inter-individual variation in WM metabolism is relevant for cognitive resilience, is associated with modifiable protective and risk factors, and may improve prediction of future cognitive decline beyond established imaging biomarkers.

Results

Summary of major analytic steps

To address our hypotheses, we conducted a structured series of analyses across the Mayo Clinic Study of Aging (MCSA) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Figure 1 provides an overview of the study framework. First, distinct WM metabolic signatures associated with baseline cognitive performance were identified in cognitively unimpaired (CU) participants from the MCSA cohort using voxel-wise FDG-PET analyses. These signatures were then validated in participants with mild cognitive impairment (MCI) from MCSA and in the full ADNI cohort, and we further tested whether WM metabolism provides unique contributions to cognition beyond traditional GM metabolism and amyloid pathology. Second, we explored the associations between WM metabolic signatures and a comprehensive set of protective and risk factors in the MCSA cohort, including demographics, genetics, systemic vascular health, brain pathologies, and WM integrity. Further, a structural equation model (SEM) was constructed to evaluate the potential driving mechanisms underlying WM metabolic signatures. Finally, we assessed the predictive value of baseline WM metabolic signatures for future cognitive decline using longitudinal data, incorporating over 15,000 follow-up observations from MCSA and ADNI combined.

Fig. 1. Overview of study.

This study addresses three questions on WM glucose metabolism in aging and dementia. 1) Cognitive related WM metabolic signatures were identified in MCSA-CU, validated in MCSA-MCI and the full ADNI cohort, and tested for contributions to cognition beyond GM-FDG and amyloid pathology. 2) Associations of WM metabolic signatures with protective and risk factors were examined in both cohorts, with potential mechanisms assessed using structural equation modeling. 3) Baseline WM metabolic levels were evaluated for their ability to predict future cognitive decline using longitudinal data from MCSA and ADNI. The glass brain rendering was generated using Surf Ice (https://www.nitrc.org/projects/surfice/) based on the results of this study. The pathological schematic was created using Figdraw.com. # means available subsets of the full cohorts. Tau-PET: MCSA, n = 499; ADNI, n = 235. WMH: MCSA, n = 1948; ADNI, n = 1050. Plasma GFAP and NfL: MCSA, n = 1696; ADNI, n = 215. Diffusion MRI neurite density was available only in MCSA (n = 259). Missing data are detailed in Supplementary Table 1. MCSA Mayo Clinic Study of Aging, ADNI Alzheimer’s Disease Neuroimaging Initiative, GM gray matter, WM white matter, CU cognitively unimpaired, MCI mild cognitive impairment, EWM expected WM, AWM atypical WM, WMH white matter hyperintensities, GFAP glial fibrillary acidic protein, NfL neurofilament light chain, SUVR standardized uptake value ratio.

Study participants

A total of 3142 participants were included, comprising 1989 individuals from the MCSA cohort (mean age 71.4 ± 10.1 years, 52.8% males) and 1153 from the ADNI cohort (mean age 72.8 ± 7.5 years, 52.3% males). Of these, 2777 had longitudinal follow-up with 15,287 total observations. The MCSA cohort contributed 1722 participants with 9948 visits (mean ± SD follow-up: 6.3 ± 3.7 years; median 6 visits, IQR: 3–8), while the ADNI cohort included 1,055 participants with 5339 visits (mean ± SD follow-up: 4.3 ± 3.1 years; median 5 visits, IQR: 3–7). In MCSA, the median interval from clinical visit to magnetic resonance imaging (MRI) was 63 days (IQR: 44–85), with FDG-PET acquired a median of 12 days after MRI (IQR: 4–28), Aβ-PET 10 days (IQR: 2–27), and Tau-PET 3 days (IQR: 0–9). In ADNI, the corresponding MRI interval was 17 days (IQR: 11–27), with FDG-PET acquired 24 days after MRI (IQR: 15–41), Aβ-PET 27 days (IQR: 17–44), and Tau-PET 40 days (IQR: 24–72). Overall, multimodal data was collected within closely aligned time windows across cohorts. Patterns of missing data across key variables in the MCSA and ADNI cohorts are summarized in Supplementary Table 1.

The baseline characteristics of the participants are summarized in Table 1. Across both cohorts, CU had a lower prevalence of APOE-ε4 carriers, fewer AD- and vascular-related pathological biomarkers, and higher educational attainment compared with those with MCI and dementia. A major distinction between the two cohorts reflects their design and resulting profiles. The MCSA is a population-based study that represents typical community older adults who exhibit both vascular comorbidities and AD pathology. Participants in MCSA had higher prevalence of hypertension, diabetes, and dyslipidemia, captured by a composite cardiovascular metabolic conditions (CMC) score. In contrast, ADNI was specifically designed to track AD progression for clinical trials, and participants have lower systemic vascular risk due to the recruitment mechanism.

Table 1.

Baseline demographic and clinical characteristics of participants

		MCSA			ADNI
	All (n = 1989)	CU (n = 1754)	MCI (n = 235)	All (n = 1153)	CU (n = 309)	MCI (n = 651)	Dementia (n = 193)
Demographics
Age, years	71.4 (10.1)	70.5 (9.9)	78.7 (8.6)	72.8 (7.5)	73.3 (6.4)	72.1 (7.6)	74.6 (8.3)
Male, no. (%)	1051 (52.8)	907 (51.7)	144 (61.3)	603 (52.3)	139 (45.0)	354 (54.4)	110 (57.0)
APOE-ε4 positive, no. (%)	564 (28.4)	476 (27.1)	88 (37.4)	515 (44.7)	88 (28.5)	299 (46.0)	128 (66.3)
Education, years	14.8 (2.6)	14.9 (2.5)	13.6 (3.1)	16.2 (2.6)	16.6 (2.6)	16.2 (2.6)	15.5 (2.6)
Vascular Risk
CMC score, median (IQR)	2 (1–3)	2 (1–3)	3 (1–4)	NA	NA	NA	NA
VRF score, median (IQR)	2 (1–2)	2 (1–2)	2 (1–3)	1 (0–1)	1 (0–1)	1 (0–1)	1 (0–1)
Hypertension, no. (%)	1047 (52.6)	881 (50.2)	166 (70.6)	562 (48.7)	148 (47.9)	321 (49.3)	93 (48.2)
Diabetes, no. (%)	710 (35.7)	607 (34.6)	103 (43.8)	102 (8.8)	26 (8.4)	60 (9.2)	16 (8.3)
Dyslipidemia, no. (%)	1218 (61.2)	1059 (60.4)	159 (67.7)	254 (22.0)	62 (20.1)	159 (24.4)	33 (17.1)
Stroke, no. (%)	113 (5.7)	85 (4.8)	28 (11.9)	18 (1.6)	3 (1.0)	10 (1.5)	5 (2.6)
Cognition
MMSE	28.2 (1.6)	28.6 (1.2)	25.6 (2.0)	27.4 (2.7)	29.0 (1.3)	28.0 (1.8)	23.0 (2.5)
Global cognition (z-score)	0.04 (1.13)	0.26 (0.94)	-1.76 (0.95)	NA	NA	NA	NA
ADAS-Cog-11	NA	NA	NA	10.18 (6.77)	5.65 (2.94)	9.24 (4.29)	20.63 (7.27)
Brain Pathology Biomarkers
Aβ positive, no. (%)	623 (31.3)	488 (27.8)	135 (57.4)	587 (50.9)	90 (29.1)	333 (51.2)	164 (85.0)
Tau positive, no. (%)	53 (2.7)	40 (2.3)	13 (5.5)	115 (10.0)	0 (0)	78 (12.0)	37 (19.2)
Vascular positive, no. (%)	401 (20.2)	309 (17.9)	92 (41.1)	200 (17.3)	48 (15.5)	101 (15.5)	51 (26.4)
Aβ Centiloid	26.8 (33.1)	23.3 (28.6)	52.8 (48.9)	40.4 (48.3)	18.6 (35.8)	39.3 (46.9)	79.2 (46.8)
Tau SUVR#	1.2 (0.1)	1.2 (0.1)	1.3 (0.2)	1.4 (0.3)	1.2 (0.1)	1.3 (0.3)	1.6 (0.4)
WMH/TIV (%)	0.87 (0.94)	0.80 (0.87)	1.41 (1.23)	0.78 (0.89)	0.74 (0.92)	0.72 (0.76)	1.05 (1.16)
Plasma GFAP, pg/ml#	113.0 (71.6)	107.2 (62.0)	158.0 (113.7)	148.6 (100.6)	138.8 (87.5)	142.9 (79.5)	368.9 (254.4)
Plasma NfL, pg/ml#	27.1 (21.2)	25.4 (19.4)	39.7 (29.1)	19.5 (10.1)	18.3 (7.6)	19.4 (10.6)	35.9 (14.9)

Data are shown as mean (SD) unless otherwise stated. CMC is a composite score based on seven conditions: hypertension, hyperlipidemia, arrhythmias, coronary artery disease, heart failure, diabetes, and stroke. VRF is a composite score based on four diseases: hypertension, hyperlipidemia, diabetes, and stroke. V+ (vascular positive) was defined as WMH/TIV (%) > 1.3%.

^# means subsets of the full cohorts. For Tau-SUVR: MCSA (CU n = 453, MCI n = 47); ADNI (CU n = 6, MCI n = 177, dementia n = 52). For plasma data: MCSA (CU n = 1503, MCI n = 193); ADNI (CU n = 88, MCI n = 120, dementia n = 7).

CU cognitively unimpaired, MCI mild cognitive impairment, CMC cardiovascular metabolic conditions, VRF vascular risk factors, IQR interquartile range, MMSE Mini-Mental State Examination, ADAS-Cog Alzheimer’s Disease Assessment Scale-Cognitive Subscale, SUVR standardized uptake value ratio, WMH white matter hyperintensity, TIV total intracranial volume, GFAP glial fibrillary acidic protein, NfL neurofilament light chain, NA not applicable.

Expected positive and atypical negative associations between cognition and WM-FDG uptake: EWM and AWM regions

In this study, “cognitive resilience” referred to individuals who maintained normal cognition despite aging, pathology, or other risk factors for cognitive decline^6,7. We conducted voxel-wise multiple regression analysis on FDG-PET data from cross-sectional MCSA-CU participants (n = 1754) to identify brain regions where glucose metabolism is associated with global cognition, adjusting for age, sex, and APOE-ε4. Multiple comparison corrections were applied using both a conservative threshold (false discovery rate [FDR], voxel-wise p < 0.05, cluster size k > 1000) and a liberal threshold (uncorrected voxel-wise p < 0.05, cluster size k > 1000) to refine statistical maps.

Under the conservative threshold, as expected, global cognition was positively associated with FDG-PET standardized uptake value ratio (SUVR) in nearly all GM areas (Fig. 2a). In WM, significant positive associations were observed mainly in the splenium and body of the corpus callosum and bilateral cingulum adjoining the cingulate gyrus, defined as expected WM (EWM; Fig. 2b). Unexpected negative associations, primarily in the anterior and superior corona radiata, were defined as atypical WM (AWM; Fig. 2c). At the liberal threshold, additional positively associated regions appeared in the posterior thalamic radiation and juxtacortical area, while negative associations extended to deep and subcortical WM, including the anterior and superior corona radiata, precentral WM, and superior longitudinal fasciculus. Peak coordinates and cluster details are provided in Supplementary Tables 2–3. To further explore these divergent FDG-cognition associations in WM, we extracted median SUVR values from FDR-corrected EWM and AWM region masks.

Fig. 2. Identification and cross-cohort validation of WM metabolic signatures in MCSA and ADNI.

In MCSA-CU participants (n = 1754), voxel-wise multiple regression between FDG-PET and global cognition was run within an explicit GM mask (a) and a WM probability mask (b, c), adjusting for age, sex, and APOE-ε4 status. Statistical maps were displayed at two thresholds: a conservative threshold (voxel-wise FDR-corrected p < 0.05, k > 1000) and a liberal threshold (uncorrected voxel-wise p < 0.05, k > 1000). Panels b and c illustrate WM regions showing positive (“Expected”, EWM-FDG) and negative (“Atypical”, AWM-FDG) associations with cognition, respectively. d, e Linear trends of FDG-PET SUVR in EWM and AWM regions, along with cognitive scores, across cohorts and clinical subgroups. Participants were ranked by subject-specific WM metabolic differences (EWM - AWM). Lines in the upper panels represent linear regression fits (two-sided). In the lower panels, Pearson correlation coefficients (r) and corresponding two-sided p values for the association between WM metabolic difference and cognition are shown; shaded areas indicate 95% CI. Sample sizes: MCSA (CU, n = 1754; MCI, n = 235) and ADNI (CU, n = 309; MCI, n = 651; dementia, n = 193). Source data are provided as a Source Data file. GM gray matter, WM white matter, CU cognitively unimpaired, MCI mild cognitive impairment, EWM expected WM, AWM atypical WM, SUVR standardized uptake value ratio.

Considering both EWM and AWM metabolism jointly contribute to cognition, we further ranked participants in each subgroup according to the EWM-AWM difference (EWM-FDG minus AWM-FDG) to examine whether this metabolic pattern can be reproduced in participants not used for WM signature discovery (MCSA-MCI and the three ADNI subgroups). Across these groups, cognition declined along the EWM-AWM gradient, although the slopes varied by clinical stage. The decline was most significant in MCSA-MCI and ADNI-dementia groups, while the association was relatively flat in ADNI-CU group. These findings suggest that the relative higher EWM and lower AWM metabolism reliably correlates with cognitive performance across cohorts and disease stages, although with variations in effect strength. (Fig. 2d, e).

WM metabolism provides information comparable to GM-FDG and Aβ-PET for cognition

We investigated the association of WM-FDG, GM-FDG (both EWM and AWM), and Aβ-centiloid to global cognition through multiple regression models that included age, sex, APOE-ε4, and education as covariates (Fig. 3). In the MCSA cohort, GM-FDG (adjusted R² = 43%), WM-FDG (44%), and Aβ-centiloid (43%) explained similar variance in cognition, and the combined model increased explanatory power to 46%. Shapley value (LMG) decomposition of R² assessed the relative importance of each predictor in regression³⁰, which confirmed that WM-FDG, GM-FDG, and Aβ-centiloid each contributed substantial, partly nonoverlapping information to global cognition, with WM-FDG and GM-FDG accounting for the largest share. In the ADNI cohort, GM-FDG, WM-FDG, and Aβ-centiloid explained 28%, 28%, and 23% of variance, respectively, and the combined model reached 37%. Here in ADNI, Shapley value again showed complementary contributions, although the effect of AWM-FDG on cognition was small despite being significant. Overall, across both cohorts, WM metabolism provides information about cognition comparable to GM-FDG and Aβ-PET, and its contribution is complementary rather than redundant.

Fig. 3. Forest plots summarize linear multiple regression models predicting baseline global cognition in the MCSA.

Panel a and ADNI b cohorts. In MCSA, global cognition was measured using global cognitive z-scores, whereas in ADNI it was assessed using the ADAS-Cog-11 total score (inverted so that higher values indicate better performance). In the top panels, points represent standardized regression coefficients (β), and error bars indicate 95% CI. Bottom panels display LMG (Shapley) relative importance values (point estimates) with bootstrap 95% CI (1000 resamples); LMG values are relative and sum to 100% within each model. All models were adjusted for age, sex, education, and APOE-ε4. Source data are provided as a Source Data file. GM gray matter, WM white matter, EWM expected WM, AWM atypical WM.

Factors associated with EWM and AWM metabolism

Given that demographic characteristics, genetics, vascular health, and mixed brain pathologies are known to be associated with GM metabolism in aging^31–36, we examined whether these same factors also contribute to inter-individual variation in WM FDG metabolism across cohorts. Age showed opposite associations with WM metabolism, being negatively correlated with EWM uptake (MCSA: r = –0.51, p = 5.3 × 10⁻¹³²; ADNI: r = –0.34, p = 5.1 × 10⁻³⁷) but positively with AWM (MCSA: r = 0.39, p = 8.7 × 10⁻⁷³; ADNI: r = 0.12, p = 1.9 × 10⁻⁵). Females consistently exhibited higher EWM uptake than males in both cohorts (MCSA p = 5.1 × 10⁻³⁵; ADNI p = 4.4 × 10⁻¹⁶), while no significant sex difference was observed for AWM (MCSA p = 0.45; ADNI p = 0.06). Vascular risk showed cohort-specific patterns: in MCSA, high vascular risk was associated with lower EWM and higher AWM uptake (EWM p = 5.4 × 10⁻²⁹; AWM p = 4 × 10⁻³⁹), whereas in ADNI the association was restricted to higher AWM uptake (p = 1.8 × 10⁻⁴) with no corresponding difference in EWM uptake (p = 0.89). ApoE-ε4 carriers consistently showed lower FDG uptake than non-carriers in both regions across cohorts. Education effects were modest and cohort-specific, observed only in MCSA, where higher education was linked to lower AWM uptake (r = –0.12, p = 3.7 × 10⁻⁸), with no significant associations for EWM or in ADNI. These results are illustrated in Fig. 4 (age, sex, and vascular risk) and Supplementary Fig. 1 (education and ApoE-ε4 status).

Fig. 4. Cross-cohort bivariate associations of WM metabolic signatures in MCSA (left) and ADNI (right).

Panels are organized by WM signatures (top, EWM; bottom, AWM) and covariates including age (a1/a2, g1/g2), sex (b1/b2, h1/h2), vascular risk (c1/c2, i1/i2; CMC in MCSA, VRF in ADNI), Aβ (d1/d2, j1/j2), WMH (e1/e2, k1/k2), and Tau (f1/f2, l1/l2). In scatter panels, lines represent fitted values from two-sided linear regression models, and shaded areas indicate 95% CI of the regression estimate. Age panels display unadjusted fits. Aβ, WMH, and Tau panels display regression models adjusted for age, sex, and APOE-ε4. Reported r values correspond to two-sided Pearson correlations; *r denotes partial correlations. For sex and vascular risk comparisons, two-sided independent-samples t-tests were performed. Violin plots represent kernel density distributions; inset boxplots show the median (center), interquartile range (box), and whiskers extending to 1.5 IQR. For Aβ, WMH, and Tau, red points indicate individuals in the abnormal biomarker range. CMC comprises hypertension, hyperlipidemia, arrhythmias, coronary artery disease, heart failure, diabetes, and stroke. VRF includes hypertension, hyperlipidemia, diabetes, and stroke. High- and low-risk categories were defined by a cohort-specific median split. Sample sizes: MCSA-All (N = 1989) and ADNI-All (N = 1153). Each panel includes all available participants, with missing data only for WMH (MCSA n = 1947; ADNI n = 1150) and Tau (MCSA n = 499; ADNI n = 235). Source data are provided as a Source Data file. EWM expected WM, AWM atypical WM, WMH white matter hyperintensities, TIV total intracranial volume, SUVR standardized uptake value ratio, CMC cardiovascular metabolic conditions, VRF vascular risk factors, IQR interquartile range.

To examine associations with brain pathology, we included five biomarkers corresponding to the revised AD biomarker framework¹: Aβ (Aβ-PET), tau (Tau-PET), axonal injury (plasma neurofilament light chain, NfL), neuroinflammation (plasma glial fibrillary acidic protein, GFAP), and vascular injury (WMH). After adjusting for age, sex, and ApoE-ε4 status, higher Aβ, Tau, and WMH burden were consistently associated with lower FDG uptake in EWM across cohorts (MCSA: Aβ r = –0.13, p = 7.4 × 10⁻⁹; WMH r = –0.22, p = 1.6 × 10⁻²²; Tau r = -0.1, p = 0.029; ADNI: Aβ r = –0.27, p = 4.2 × 10⁻²¹; WMH r = –0.17, p = 1.1 × 10⁻⁸; Tau r = -0.28, p = 1.7 × 10⁻⁵; Fig. 3d–f). In contrast, AWM showed weaker and more selective associations, with significant associations observed only for Aβ (MCSA: r = –0.05, p = 0.032; ADNI r = -0.06, p = 0.034), whereas WMH and tau were not significantly associated after covariate adjustment (Fig. 3j–l). Plasma biomarkers showed broader associations, with both EWM and AWM uptake negatively associated with GFAP and NfL in MCSA and ADNI (Supplementary Fig. 1). These findings suggest that EWM metabolism is broadly associated with aging, vascular burden, and multiple neuropathological markers, while AWM metabolism shows weaker and more selective pathological associations.

Contrasting correlations of EWM and AWM metabolism

To further characterize the framework of associations linking demographic, vascular, genetic, and pathological factors to WM metabolism and cognition, we fitted SEM separately in the MCSA and ADNI cohorts (Fig. 5). Both models showed good fit, and full pathways and effect estimates are provided in Supplementary Tables 4–7. Across cohorts, SEMs revealed that demographic, genetic, vascular, and pathological factors jointly shaped EWM and AWM metabolism, which in turn contributed to cognition in opposite directions. Specifically: (i) Age and vascular risk showed bidirectional associations with WM metabolism, with older age associated with lower EWM but higher AWM uptake (β_total: EWM –0.48/–0.31; AWM 0.32/0.1 in MCSA/ADNI), and a similar pattern observed for vascular risk (β_total: EWM –0.05/–0.01; AWM 0.21/0.1); (ii) Male sex consistently associated with lower EWM metabolism (β_total: –0.22/–0.2), while only minor indirect effects on AWM; (iii) APOE-ε4 exerted indirect effects on WM metabolism through its association with amyloid burden; and (iv) pathological associations differed by WM regions, with amyloid burden negatively associated with both EWM and AWM across cohorts (EWM β_total: –0.12/–0.30; AWM β_total: –0.06/–0.09), whereas WMH showed consistently negative associations with EWM (β_total: –0.18/–0.13) but minimal associations with AWM. Decomposition of total effects suggested that aging and vascular risk were the dominant positive contributors to AWM metabolism, whereas amyloid burden demonstrated an opposing association. In contrast, EWM metabolism was associated with multiple demographic and pathological factors, including age, sex, APOE-ε4, amyloid, and WMH, with relatively weaker contributions from vascular risk (Fig. 5b, d).

Fig. 5. Structural equation modeling (SEM) of age, sex, APOE-ε4, vascular diseases, and brain pathologies on WM metabolism and cognition across cohorts.

a SEM of the MCSA cohort [n = 1923]. Red arrows represent positive associations, and blue arrows represent negative associations. The standardized path coefficients were displayed above the arrows. ***p < 0.001, **p < 0.01, *p < 0.05. b Standardized direct, indirect, and total effects coefficients on WM metabolism were evaluated in the MCSA model, both separately and as a summed effect. Direct effects represent the primary path from predictor to outcome, indirect effects reflect mediation through intermediate variables (such as amyloid or WMH), and total effects are the sum of direct and indirect pathways. c SEM of the ADNI cohort [n = 1150]. d Standardized direct, indirect, and total effects coefficients on WM metabolism of the ADNI model. Vascular risk in both SEM models was indexed using the VRF score, a composite score based on four diseases: hypertension, hyperlipidemia, diabetes, and stroke. In MCSA, global cognition was measured using global cognitive z-scores, whereas in ADNI it was assessed using the ADAS-Cog-11 total score (inverted). CFI > 0.95, TLI > 0.95, RMSEA < 0.05, and SRMR < 0.05 indicate an excellent fit. Source data are provided as a Source Data file. WM white matter, EWM expected WM, AWM atypical WM, WMH white matter hyperintensities, SUVR standardized uptake value ratio, CFI comparative fit index, TLI Tucker–Lewis index, RMSEA root mean square error of approximation, SRMR standardized root mean square residual.

The contribution and predictive role of EWM and AWM metabolism in global cognition across cohorts

We conducted longitudinal analyses using linear mixed-effects model to examine whether baseline WM metabolism predicted subsequent cognitive decline. Models included baseline age, sex, education, APOE-ε4, WM metabolism, Aβ, WMH burden, follow-up time, visit cycle, and their interactions with time. Across both MCSA and ADNI, significant WM metabolism × time interactions indicated that baseline WM-FDG levels were associated with differences in the rate of cognitive decline (Table 2). Specifically, higher EWM-FDG and lower AWM-FDG were consistently associated with a slower rate of cognitive decline, with a stronger effect for EWM than AWM across cohorts (Fig. 6b). Among other variables, ApoE-ε4 and Aβ also interacted with time, with ApoE-ε4 carriers and those with higher baseline Aβ deposition showing faster cognitive decline. Increased WMH burden was associated with accelerated cognitive decline in MCSA (p = 7.2 × 10^–9) but not in ADNI (p = 0.06).

Table 2.

Mixed effect models predicting longitudinal global cognitive changes

	MCSA (R2 = 0.51)		ADNI (R2 = 0.36)
	βstd (95% CI)	p	βstd (95% CI)	p
Intercept	0.194 (0.129, 0.259)	5.6 × 10–9	–0.067 (–0.154, 0.021)	0.135
Time from baseline, years	–0.394 (–0.448, –0.339)	3.2 × 10-44	–0.209 (–0.263, -0.155)	2.8 × 10–14
Number of visits	0.285 (0.222, 0.347)	6.4 × 10–19	–0.033 (–0.082, 0.017)	0.195
Demographic variables
Baseline age, years	–0.445 (–0.501, –0.389)	4.1 × 10-51	0.063 (0.006, 0.120)	0.029
Males	–0.171 (–0.256, –0.086)	7.7 × 10-5	–0.020 (−0.126, 0.086)	0.715
APOE-ε4 carrier	–0.078 (-0.170, 0.013)	0.094	−0.138 (–0.250, –0.026)	0.016
Education, years	0.346 (0.306, 0.386)	2.1 × 10–60	0.109 (0.059, 0.160)	2.4 × 10–5
Baseline age × time	-0.093 (-0.105, -0.082)	1.0 × 10–55	0.000 (–0.018, 0.018)	0.977
Males × time	0.005 (–0.012, 0.021)	0.557	0.059 (0.027, 0.090)	2.5 × 10–4
APOE-ε4 carrier × time	–0.027 (–0.045, –0.009)	0.003	–0.048 (–0.081, –0.015)	0.005
Education × time	-0.005 (-0.013, 0.003)	0.193	0.000 (–0.015, 0.015)	0.996
WM metabolism
EWM FDG SUVR	0.268 (0.218, 0.317)	1.7 × 10–25	0.474 (0.413, 0.535)	1.6 × 10–47
AWM FDG SUVR	–0.204 (–0.249, –0.160)	6.1 × 10–19	–0.130 (–0.185, –0.076)	3.4 × 10–6
EWM FDG SUVR × time	0.048 (0.037, 0.059)	2.5 × 10–18	0.087 (0.068, 0.107)	3.2 × 10–18
AWM FDG SUVR × time	–0.027 (-0.036, –0.017)	7.0 × 10–7	–0.019 (–0.036, –0.002)	0.026
Pathologies
Aβ Centiloid	–0.251 (–0.292, –0.209)	1.5 × 10–31	–0.445 (–0.500, -0.389)	2.5 × 10–50
WMH/TIV (%)	–0.082 (–0.125, –0.038)	2.2 × 10–4	0.019 (–0.033, 0.071)	0.474
Aβ Centiloid × time	–0.102 (–0.112, –0.092)	4.8 × 10–92	–0.200 (−0.222, −0.181)	1.7 × 10–78
WMH/TIV (%) × time	–0.030 (–0.040, –0.020)	7.2 × 10–9	0.020 (–0.001, 0.035)	0.060

Linear mixed-effects models were fitted separately in MCSA and ADNI. Two-sided tests were used for all analyses. R² for fixed effects, standardized coefficients (β_std), 95% CIs, and exact p values are reported. In MCSA, global cognition was measured using global cognitive z-scores (higher scores indicate better cognition). In ADNI, cognition was assessed using the ADAS-Cog-11 total score (higher scores indicate poorer cognition), which was inverted prior to modeling to align the direction of effects across cohorts. MCSA (n = 1722 participants; 9948 observations) and ADNI (n = 1055 participants; 5339 observations).

EWM expected white matter, AWM atypical white matter, SUVR standardized uptake value ratio, ADAS-Cog Alzheimer’s Disease Assessment Scale-Cognitive Subscale, WMH white matter hyperintensity, TIV total intracranial volume.

Fig. 6. Predictive role of WM metabolism in longitudinal cognition and its relationship with WM integrity.

a Distribution of FDG-PET SUVR in EWM and AWM among cognitively unimpaired participants from MCSA and ADNI, with cut-offs defined at the 25th (red) and 75th (blue) percentiles. b Longitudinal trajectories of cognition stratified by baseline WM metabolism. Shaded colored bands indicate 95% CI of the estimated mean. Outcomes are predicted global z-score (MCSA) and sign-reversed ADAS-Cog-11 (higher values indicate better cognition) in ADNI. Higher EWM-FDG ( > 75th percentile) was consistently associated with slower cognitive decline, whereas higher AWM-FDG ( < 25th percentile) predicted faster decline in both MCSA and ADNI. Stratification by Aβ status further showed that these effects persisted in both Aβ– and Aβ+ groups, with the impact of EWM being amplified in Aβ+ individuals. c Associations between WM integrity and WM metabolism in a subset of MCSA-CU participants with NODDI data (n = 235). Higher neurite density index (NDI) was positively associated with EWM-FDG and negatively associated with AWM-FDG. Asterisks (*) denote partial correlation coefficients (two-sided) adjusted for age, sex, and APOE-ε4 status, and shaded areas indicate 95% CI. Source data are provided as a Source Data file. WM white matter, EWM expected WM, AWM atypical WM, SUVR standardized uptake value ratio, NDI neurite density index, SCC splenium of the corpus callosum, CGC cingulum adjoining the cingulate gyrus, ACR anterior corona radiata, SCR superior corona radiata.

Considering the interaction between Aβ pathology and time, we examined cognitive trajectories stratified by Aβ status and observed a graded pattern of cognitive decline. Figure 6b depicts longitudinal cognitive trajectories with slopes reflecting the rate of cognitive decline as a function of baseline FDG levels. When stratified by Aβ status, the protective effect of high EWM metabolism and the harmful effect of high AWM metabolism persisted in both Aβ– and Aβ+ groups. Notably, the effect of EWM was markedly amplified in the Aβ+ group compared with AWM. Baseline intercept differences reflect baseline cognitive performance and were accompanied by differences in age, vascular risk, and pathological burden (Supplementary Tables 8–9). All these results indicate that WM metabolism provides information on longitudinal cognitive change beyond established risk factors. Higher EWM and lower AWM metabolism at baseline were associated with more favorable cognitive trajectories, whereas lower EWM and higher AWM metabolism were associated with faster decline.

Preserved WM integrity supports healthy WM metabolism

We further analyzed regional neurite density index (NDI) in a subset of MCSA-CU participants (n = 235) who had neurite orientation dispersion and density imaging (NODDI) data (Fig. 6c). Significant positive correlations were observed between regional NDI in the splenium of the corpus callosum (SCC: r = 0.31, p = 1.1 × 10^–6) and cingulum adjacent to the cingulate gyrus (CGC: r = 0.21, p = 0.001) with FDG-PET uptake in EWM. In contrast, regional NDI in the anterior and superior corona radiata (ACR: r = –0.16, p = 0.012; SCR: r = –0.18, p = 0.006) showed negative correlations with FDG-PET uptake in AWM. All these associations are adjusted for age, sex, and ApoE-ε4 status. Because higher NDI values reflect healthier WM microstructural integrity, these results suggest that the link between higher EWM metabolism and lower AWM metabolism with better cognitive performance may be supported by preserved WM integrity.

Robustness of WM metabolic signatures

To evaluate the robustness of WM metabolic signatures identified in the MCSA-CU voxel-wise regressions, we conducted a series of sensitivity analyses. At the image-processing level, we compared partial volume corrected versus uncorrected images to evaluate partial volume effects and applied different Gaussian smoothing kernels (4, 6, and 8 mm) to examine potential gray matter spillover. At the modeling-level, we tested adding covariates (age, sex, APOE-ε4), further adjusting for brain parenchymal fraction [total tissue volume (TTV)/total intracranial volume (TIV)], and excluding participants with severe atrophy (TTV/TIV <mean-1.5 SD by age- and sex-stratified distributions). These variations did not materially alter the spatial distribution of EWM and AWM clusters or their associations with cognition (see Supplementary Figs. 2–4), indicating that the observed associations were not attributable to image-processing noises or confounding from demographics or brain atrophy. In addition, we repeated the cross-sectional regressions and longitudinal mixed-effect models using Mini-Mental State Examination (MMSE) as the outcome variable, with results provided in the Supplementary Tables 10–11, which were consistent with the main analyses.

Discussion

We investigated WM metabolism in aging and dementia, an area less studied than GM, using multimodal neuroimaging data across MCSA and ADNI cohorts. We identified two divergent patterns of WM metabolism: higher glucose metabolism in EWM (corpus callosum and cingulum) was linked to better cognition, whereas higher glucose metabolism in AWM (corona radiata) was linked to worse cognition. Cross-sectionally, EWM metabolism was lower in individuals with greater age, Aβ burden, and WMH, whereas higher AWM metabolism was observed in people with older age and vascular risk, with reduced levels in the presence of greater AD pathology (Aβ and APOE-ε4). Longitudinally, individuals with higher EWM and lower AWM metabolism at baseline exhibited greater resilience against future cognitive deterioration. These findings suggest that WM metabolism undergoes distinct and, in some regions, opposite patterns of metabolic associations compared with GM, and that both EWM and AWM provide important and complementary information for understanding cognitive aging and dementia-related vulnerability.

Since the invention of ¹⁸F-FDG as a neurochemical tracer in the 1970s, ¹⁸F-FDG PET has become widely utilized in both research and clinical practice of neurodegenerative diseases globally. There is a well-established consensus recognizing GM hypometabolism as a characteristic feature of aging and dementia, with distinct regional vulnerabilities: the temporoparietal cortex and cingulate gyrus in AD¹⁷, and the frontal lobe in cerebrovascular disease^18,23. Consistent with prior studies and corroborated by our findings, hypometabolism in GM is significantly related to aging and cognitive decline^17,24. In comparison, relatively few studies have examined WM metabolism using FDG-PET, and most prior work has reported WM hypometabolism. For example, hippocampal atrophy has been shown to precede glucose hypometabolism in the cingulate gyrus and subgenual cortices²⁵. Reduced glucose metabolism in the cingulum of AD patients has also been reported²². Similarly, reduced FDG uptake in WMH regions of Aβ+ compared to Aβ- individuals has been observed²⁶. These results align with our findings that reduced FDG uptake in the cingulum and corpus callosum (EWM) is associated with worse cognition, consistent with prior studies linking AD-related hypometabolism to corpus callosum atrophy and microstructural damage^27,28.

In contrast to the predominant evidence of hypometabolism, some studies have reported opposite findings, suggesting that regional WM metabolism may also increase. A 2017 retrospective pilot study of 18 AD patients and 18 healthy controls first reported hypermetabolism in frontal and parietal WM, which was negatively correlated with MMSE scores²⁹. More recently, a multi-tracer PET study in patients with cerebral small vessel disease reported elevated glucose metabolism and glycolysis in non-lesional WM compared to WMH regions¹⁵, although no direct relationship with cognition was established. Although notable, both studies have received limited attention due to small sample sizes, lack of pathological confirmation, and absence of longitudinal validation. In the present work, we observed increased WM metabolism in the corona radiata, which was unexpectedly associated with worse cognition; under a liberal statistical threshold, these regions extended to precentral WM and superior longitudinal fasciculus.

By including more than 3000 individuals with over 15,000 visits across two independent cohorts and more than a decade of follow-up, our study provides substantially greater statistical power and reproducibility. This allowed us to consistently demonstrate decreased glucose metabolism in the cingulum and corpus callosum and relative increases in deep WM, linking these spatially divergent patterns to cognition, aging, and vascular risk. Our results confirm previously reported phenomenon of WM metabolic heterogeneity and examine metabolic heterogeneity of WM in two large, well-characterized cohorts.

WM tracts contribute to cognition by serving as the major highways for information transfer across the cerebral cortex. The corpus callosum is the largest commissural fiber, connecting homologous cortical regions across both hemispheres, and is essential for interhemispheric integration, supporting functions from sensory and motor processing to higher-order cognition. Cingulum is a long association bundle running within the cingulate gyrus and parahippocampal region, interconnecting frontal, parietal, and medial temporal cortices as well as subcortical nuclei. It serves as a structural backbone of the limbic system and default mode network, thereby supporting memory, attention, and executive function³⁷. Both regions are vulnerable to age- and AD-related neurodegeneration, with preserved integrity associated with better cognitive outcomes and compensation in aging^38,39.

The corona radiata is a dense bundle of fan-shaped projection fibers radiating from the internal capsule to widespread cortical regions, carrying most neural traffic between the cerebral cortex and subcortical structures, including the thalamus, brainstem, and spinal cord. Owing to its location in deep WM supplied by small penetrating arteries, it is particularly vulnerable to microvascular pathology⁴⁰, ischemic, and neurodegeneration, making it one of the most affected regions in vascular cognitive impairment and dementia^41,42. Structural and microstructural changes in the corona radiata are strongly associated with declines in processing speed, attention, and executive function, with the superior corona radiata showing especially strong links to reduced cognitive flexibility⁴³. Nevertheless, this tract also exhibits notable plasticity, with evidence of structural preservation in successful aging^44,45 and functional reorganization following rehabilitation⁴⁶. Collectively, the unique anatomy, functions, and vulnerabilities of these WM tracts form the foundation of WM metabolism, supporting that WM metabolism changes are not uniform but regionally specific with aging or vascular pathology.

Although no prior studies have comprehensively examined relationships between WM glucose metabolism and protective or risk factors, our findings for EWM generally parallel those reported for GM metabolism. Aging was related to widespread hypometabolism, while females show higher brain metabolism than males, consistent with prior studies of age- related and sex differences in cerebral glucose metabolism^31,32. The observation that the adult female brain appears metabolically younger than the male brain³³, which also termed neoteny, has been linked to hormonal influences and sex-related differences in glial biology⁴⁷. Prior work has also reported associations with Aβ, tau, WMH, and vascular burden with regional metabolic reductions^34–36. However, the associations observed in AWM regions differed from those typically seen in GM. For AWM, age and systemic vascular risk were associated with higher FDG uptake, whereas Aβ burden and APOE-ε4 carriage were associated with lower uptake. These divergent patterns suggest that EWM and AWM may reflect distinct biological processes. Decreased EWM-FDG may be consistent with neurodegenerative or demyelination-related changes associated with aging, Aβ accumulation, and WMH burden, which known to accompany reduced axonal integrity and metabolic activity. Conversely, higher AWM-FDG could be compatible with mechanisms such as reactive glial responses, altered glycolytic activity, or less efficient energy utilization in the context of vascular stress, although current in vivo evidence in humans remains sparse. Thus, the WM components with opposing metabolic trends may reflect the coexistence of two partially dissociable biological processes: neurodegenerative hypometabolism and vascular/inflammatory-related hypermetabolism. These mechanisms are not mutually exclusive and may operate concurrently within the aging brain, but further multimodal and molecular studies will be required to determine how they interact and their relative contribution to longitudinal cognitive function.

This interpretation is also supported by literature that chronic hypoperfusion reduces oxygen and nutrient delivery, causing ischemic stress that activates astrocytes and microglia and shifts energy metabolism from oxidative phosphorylation to anaerobic glycolysis^48,49. Glial cells, particularly microglia, exhibit higher glycolytic rates and glucose uptake than neurons and astrocytes during inflammatory activation, producing elevated FDG-PET signal in WM⁵⁰. Mitochondria dynamics may also contribute to this process, as intercellular mitochondrial transfer under stress conditions supports energy homeostasis and tissue repair⁵¹.

After identifying EWM and AWM, we evaluated their explanatory utility for baseline global cognition and longitudinal cognitive decline across two independent cohorts (MCSA and ADNI). Both cross-sectional and longitudinal analyses consistently demonstrated that WM metabolism provides comparable and non-overlapping information in explaining cognition, compared to GM metabolism and Aβ, highlighting its independent value. Most aging studies have focused on decreased FDG uptake in GM, reflecting the neuronal loss and synaptic dysfunction. However, a recent mouse brain spatiotemporal RNA-seq study revealed that glia aging progresses more rapidly in WM than in cortical regions⁵², emphasizing the need to explore multiple parallel and interacting mechanisms across tissue types to map the causal evolution of aging.

Linear mixed-effects models revealed significant interactions between time and WM metabolism in relation to cognitive decline rates. Baseline higher EWM metabolism and lower AWM metabolism were associated with slower cognitive deterioration, independent of Aβ and WMH. These associations remained consistent across both cohorts, indicating the robustness of WM metabolic signatures as predictors of longitudinal cognitive trajectories beyond classical Aβ-related pathways. WM integrity findings further supported the cognitive results, as reflected by opposing correlations between NDI and glucose metabolism in EWM and AWM (Fig. 6c). In EWM, higher metabolism corresponded to better WM integrity, whereas in AWM, lower metabolism was associated with lower vascular risk and higher NDI. In this context, elevated AWM metabolism is more consistent with maladaptive metabolic states, potentially reflecting gliosis-related activity or reduced energetic efficiency rather than successful compensation. These findings help contextualize why higher EWM and lower AWM metabolism at baseline are associated with flatter cognitive decline trajectories, consistent with differences in WM microstructural integrity and network efficiency. A recent diffusion MRI study reported that cognitive resilience involves reorganization of the limbic and default mode network⁵³, with the cingulum bundle and corona radiata serving as key connectivity pathways. In this context, our findings align with the network-efficiency framework by identifying distinct WM metabolic signatures that may serve as markers of preserved and altered brain glucose metabolism.

To note, we used the pons as the reference region for SUVR normalization given its metabolic stability across aging and AD and its minimal involvement in disease pathology⁵⁴. Consequently, WM SUVR values were generally below 1.0, which is expected since WM glucose uptake is roughly half that of GM⁵⁵. This reflects the lower synaptic density and signaling activity in WM, where energy consumption primarily supports myelin and axonal maintenance rather than active neurotransmission. To ensure robustness, we conducted a series of sensitivity analyses to test the robustness of WM metabolic signatures against variations in image preprocessing and model specification, including partial volume correction (PVC), different smoothing kernels (4–8 mm), additional covariates, and brain atrophy control. The voxel-wise regression results remained consistent (Supplementary Fig. 2–4), confirming that our findings were not driven by preprocessing choices or atrophy-related artifacts.

Cohort-specific differences were observed in both cross-sectional and longitudinal analyses. Within the SEM framework, age and vascular risk were key drivers of increased AWM metabolism in both cohorts, but the total effect was smaller in ADNI. Similarly, in the longitudinal mixed-effects models, WMH significantly contributed to cognitive decline in MCSA but not in ADNI. This discrepancy likely reflects fundamental cohort characteristics: MCSA is a population-based sample with a broader age range and higher prevalence of vascular risk factors, whereas ADNI is a clinic-based and enriched for AD, with narrower age dispersion and minimized vascular comorbidities (Hachinski score >4 was an exclusion criteria)⁵⁶. As a result, both vascular risk and WMH burden was markedly lower in ADNI than in MCSA, reducing their apparent impacts on WM metabolism and cognition. Notably, the ongoing ADNI4 aims to mitigate this bias through broader recruitment of underrepresented groups and enhanced measurements of cerebrovascular disease, thereby improving generalizability and alignment with population-based studies.

Our study has important strengths, including a large-sample, longitudinal investigation of WM glucose metabolism and cognitive resilience, with concurrent consideration of amyloid, vascular, and neuroinflammatory pathologies across the MCSA and ADNI cohorts. However, several limitations should be noted. First, our measures of WM integrity (NDI), WMH, and neuroinflammation (plasma GFAP) cannot fully capture the cellular and molecular mechanisms underlying WM metabolic dysregulation. Future multimodal and experimental studies are needed to address this from a biological perspective. Second, Tau-PET data was available for only ~20% of participants, limiting formal evaluation of tau’s role in SEM analysis. Third, cohort differences (MCSA’s population-based design versus ADNI’s AD-enriched sample) and limited data availability (e.g., GFAP and NDI in ADNI, particularly after matching concurrent FDG-PET scans) restricted full cross-validation. Fourth, although identifiable cases of non-AD dementias were excluded from ADNI analysis, unrecognized co-existing pathologies such as α-synuclein or TDP-43 cannot be fully ruled out, and future studies with comprehensive neuropathologic data are needed to disentangle these sources of heterogeneity. Fifth, the spatial localization of EWM and AWM in this study was derived from a WM template instead of subject-specific tractography. Given the resolution of FDG-PET and the convergence of multiple fiber pathways in these regions, tract references in our interpretation should be considered anatomically approximations rather than precise tract-level identification. Future work using individualized diffusion tractography will be required to refine the specific tract pathways underlying these metabolic patterns. Sixth, we used global WMH rather than regional segmentation, which may obscure tract-specific effects. Finally, plasma GFAP primarily reflects astrocytic reactivity and does not capture microglial activation, oligodendroglia responses, or inflammatory signaling cascades, which could be better disentangled by emerging TSPO-PET tracers and autoradiographic approaches providing cellular-level resolution.

In conclusion, our study demonstrates that EWM and AWM represent complementary metabolic features capturing differentially weighted aspects of aging-related WM vulnerability. EWM shows greater susceptibility to AD-related degeneration, whereas AWM appears more strongly influenced by vascular-related alterations. Together, these patterns shed light on the heterogeneity of WM metabolic processes and their differential associations with cognitive decline. Elevated AWM metabolism may reflect maladaptive metabolic states, potentially related to inflammatory activity or reduced energetic efficiency, rather than successful compensation. Further investigation into the neuroinflammatory and cellular energy mechanisms underlying WM metabolism may deepen our understanding of brain resilience and inform targeted interventions for AD and cerebrovascular disease.

Methods

Selection of participants

This study included participants from two independent cohorts: MCSA, a population-based prospective study with a stratified random sampling design, aimed to investigate cognitive aging among residents of Olmsted County, Minnesota, USA⁵⁷. ADNI, a multicenter longitudinal study designed to investigate clinical, imaging, genetic, and biological markers for the early detection and tracking of MCI and AD⁵⁸. Study details are available at: MCSA: https://www.mayo.edu/research/centers-programs/alzheimers-disease-research-center/research-activities/mayo-clinic-study-aging/overview; ADNI: https://adni.loni.usc.edu/.

For cross-sectional analyses, eligible participants were ≥50 years old, had concurrent amyloid PET, FDG-PET, structural MRI scans, and completed neuropsychological assessments. Those with incomplete clinical data or uncertain cognitive diagnoses were excluded. In ADNI, we further reviewed available diagnostic and neuropathological information and excluded cases with dementia with Lewy bodies (DLB), frontotemporal lobar degeneration with TDP-43 (FTLD-TDP), or limbic-predominant age-related TDP-43 encephalopathy (LATE). For longitudinal analyses, participants were required to have ≥2 cognitive assessments. Given differences in cohort composition and diagnostic criteria, MCSA and ADNI were analyzed separately.

Sex was defined based on biological sex as recorded in the medical record at enrollment. Sex was included as a covariate in all regression, linear mixed-effects, and SEM to account for potential sex-related differences in brain metabolism and cognition. In addition, group comparisons of WM metabolic signatures were performed between males and females. All analyses were conducted at the individual participant level (biological replicates only). No technical replicates were used.

Standard protocol approvals, registrations, and patient consents

The MCSA was approved by the Mayo Clinic and Olmsted Medical Center Institutional Review Boards, and informed consent was obtained from all participants or their legally authorized representatives. ADNI was approved by the institutional review boards of all participating sites, with written informed consent from all participants.

Cognitive evaluation

In MCSA, cognition was assessed by a neuropsychological battery covering executive function, language, memory, and visuospatial performance; test scores were standardized and averaged into a global cognitive z-scores (higher scores indicate better cognition)⁵⁷. In ADNI, global cognition was measured by the Alzheimer’s Disease Assessment Scale-Cognitive Subscale (ADAS-Cog-11), with scores inverted (multiplied by –1) so that higher values reflected better performance, consistent with MCSA scaling. Because of differences in assessment protocols, cognitive-related analyses were conducted separately within each cohort. MMSE scores were also available in both cohorts and are reported but were not used in the primary analyses due to limited sensitivity among CU participants. Results using MMSE as the cognitive outcome are provided as sensitivity analyses in the Supplementary Tables 8–9.

MR acquisition and processing

For MCSA, all MRI images were acquired on 3 T scanners. Each participant underwent 3D T1-weighted MPRAGE and 3D FLAIR imaging. T1 was used for PVC⁵⁹, normalization, and tissue volumetry⁶⁰. WMH was segmented from FLAIR using a validated in-house pipeline and normalized to TIV (WMH/TIV%)⁶¹. In ADNI, MPRAGE and FLAIR images were processed with the same pipeline as MCSA. A subset of MCSA participants had multi-shell diffusion MRI enabling NODDI, with acquisition and processing details that can be found in previous publications⁶². The JHU ‘Eve’ WM atlas was warped to participant native space to derive regional NODDI measures for certain WM tracts involved in the significant FDG-cognition results. NDI was used as a marker of WM integrity in this study²⁰. NODDI data were not available in the ADNI. Full details are provided in the Supplementary Methods.

FDG-PET acquisition and processing

The acquisition and processing procedures for FDG-PET in the MCSA⁶³ and ADNI have been detailed provided in the Supplementary Methods. We analyzed FDG-PET images with an automated in-house pipeline using SPM12⁶⁴. Individual PET frames were visually inspected for motion and averaged. Each PET image was rigidly co-registered (6 DOF) to the participant’s T1 images. T1 images were segmented into GM, WM, and cerebrospinal fluid probability maps, and the MRI-derived deformation fields were applied to spatially normalize PET images to the Mayo Clinic Adult Lifespan Template (MCALT; https://www.nitrc.org/projects/mcalt/) space. For voxel-wise analyses, explicit GM and WM masks were constructed in template space from mean tissue probability maps across participants. GM voxels were defined using a probability threshold >0.20 and WM voxels using >0.50. To prevent overlapping between compartments, voxels were assigned to GM when GM > WM and WM when WM > GM. Mask distributions were visually inspected to ensure anatomical plausibility. PVC was performed using a two-compartment model⁶⁵. Images were smoothed with a 6 mm full-width at half-maximum Gaussian kernel (with sensitivity analyses at 4 and 8 mm) and normalized to the pons to generate SUVR maps. Regional SUVRs for GM, EWM, and AWM were calculated as the median uptake within masks defined from FDR-corrected voxel-wise regression results.

Amyloid and tau PET acquisition and processing

Amyloid-PET used ¹¹C-PiB in MCSA and ¹⁸F-florbetapir or ¹⁸F-florbetaben in ADNI. A subset of participants from both cohorts underwent tau-PET with ¹⁸F-flortaucipir (MCSA: n = 499 [25%], ADNI: n = 235 [20%]). Acquisition details are provided in the Supplementary Methods. Amyloid and tau PET images were motion-checked, frame-averaged, rigidly co-registered (6 DOF) to the participant’s T1-weighted MRI and spatially normalized to the Mayo MCALT atlas using MRI-derived deformation fields. Global amyloid-PET SUVR was calculated as the voxel-count–weighted median across predefined cortical regions (prefrontal, orbitofrontal, parietal, temporal, anterior cingulate, and posterior cingulate/precuneus), scaled to cerebellar crus GM, and converted to centiloid units. For ADNI, centiloid values processed by UC Berkeley were downloaded from the LONI database. Global tau-PET SUVR was calculated as the median uptake across entorhinal, amygdala, parahippocampal, fusiform, inferior temporal, and middle temporal regions, also scaled to cerebellar crus GM.

Plasma biomarkers

Both MCSA and ADNI participants had previously collected fasting blood samples, which were centrifuged and stored at –80 °C until testing. In MCSA, plasma GFAP and NfL were measured using the Simoa® Neurology 4-Plex E Advantage kit (N4PE) on a Quanterix Simoa HD-X analyzer (Quanterix, Lexington, MA)⁶⁶. In ADNI, plasma biomarkers was measured using the same platform and protocol. Availability was limited to a subset of participants in each cohort, as summarized in Supplementary Table 1.

Revised AD biomarker categorization

According to the latest modified AD biomarker categories¹, we evaluated four pathogenic pathways using imaging measures and plasma analytes: amyloid (A) pathology via amyloid-PET, tau (T) pathology via tau-PET, neurodegeneration (N) via NfL, inflammatory (I) activity via plasma GFAP, and vascular injury (V) via WMH. Specifically, A+ was classified by an established cut-off (centiloid ≥ 22)⁶⁷, T+ was defined as SUVR ≥ 1.29, and V+ was defined as WMH/TIV% >1.3%^68,69. Although inflammation is recognized as an important mechanism in AD pathogenesis⁷⁰ and plasma GFAP is emerging as a sensitive marker of astrocytic reactivity associated with early amyloid deposition, a widely accepted cut-off value for GFAP is not yet available.

APOE-ε4, education, and vascular risk

APOE-ε4 positivity was defined as carrying one or more ε4 alleles. Intellectual enrichment was assessed by years of education. In MCSA, systemic vascular health was measured by a composite CMC score based on the presence or absence of seven cardiovascular metabolic conditions: hypertension, hyperlipidemia, cardiac arrhythmias, coronary artery disease, congestive heart failure, diabetes mellitus, and stroke^71,72. As the ADNI cohort does not capture all seven conditions, we also constructed a harmonized vascular risk factor (VRF) score based on the four vascular conditions available in both cohorts (hypertension, diabetes, hyperlipidemia, and stroke). Creating this harmonized VRF score enabled consistent cross-cohort interpretation while maintaining conceptual equivalence with existing vascular burden measures. In the final analytic sample, the median VRF score was 2 in MCSA and 1 in ADNI.

Statistical analysis

All statistical analyses except for voxel-wise statistics were conducted with R 4.4.0.

Cognition-associated WM glucose metabolism regions

Voxel-wise multiple linear regression was performed on MCSA-CU participants (n = 1754) to examine associations between z-global cognition scores and FDG SUVR, with age, sex, and ApoE-ε4 genotype as covariates. GM and WM masks were applied as explicit masks in MCALT space, derived from the mean segmented and normalized probability maps across all participants. Probability thresholds of GM > 0.20 and WM > 0.50 were used. A higher threshold was chosen for WM to account for its proximity to GM and less distinct boundaries, whereas a lower threshold was applied for GM due to its clearer cortical definition and sharper tissue contrast. To avoid overlapping, voxels were assigned to GM when GM > WM and to WM when WM > GM. The resulting masks were visually inspected to ensure anatomical plausibility. Voxel-wise results were corrected for multiple comparisons using both a conservative threshold (voxel-level FDR-corrected, p < 0.05, cluster size k > 1000 mm³) and a liberal threshold (uncorrected voxel-level p < 0.001, cluster size k > 1000 mm³). Specific WM anatomical labels inside the significant clusters were reported based on the JHU “Eve” WM atlas⁷³. In addition, we extracted the FDG-PET SUVR in the “EWM and AWM regions” of each participant using ROIs mask derived from the FDR-corrected voxel-wise multiple regression maps for further analysis.

Cross-sectional contribution of WM metabolism to global cognition variance

We evaluated the contribution of WM metabolism to baseline global cognitive performance compared to the established imaging markers (GM-FDG and Aβ). Multiple linear regression models were constructed within each cohort to assess the associations between imaging measures and global cognition. Each model included demographic covariates (age, sex, and education) and APOE-ε4 status. Imaging variables (FDG SUVR in EWM, AWM, and GM, and Aβ centiloid) were first included separately in the regression model and then combined into one model. To ensure model stability, multicollinearity among predictors was evaluated using variance inflation factor (VIF), tolerance, and condition index. All predictors met conventional thresholds (VIF < 5, tolerance >0.2). Although the maximum condition index exceeded 30, no joint variance concentration was detected (i.e., no two predictors shared >50% of variance at the same index), confirming that multicollinearity did not bias model estimation. We reported the adjusted R² to evaluate the overall fit of the regression models. In addition, we calculated the Shapley value (LMG) decomposition of R² to assess the relative importance of each predictor³⁰.

Risk/protective factors and their impact on regional WM metabolism

The effects of sex, ApoE-ε4 genotype (carrier vs. non-carrier), and vascular risk (high vs. low) on WM metabolism were examined by independent samples t-test. Associations with continuous demographic variables (age and education) were tested using Pearson correlation analysis. To investigate relationships between WM metabolism and various pathological markers, we performed partial correlation analyses adjusting for age, sex, and APOE-ε4 status. In these models, WM metabolic measures (EWM and AWM FDG SUVR) and biomarker values (Aβ centiloid, tau SUVR, WMH/TIV%, plasma GFAP, and plasma NfL) were entered directly, and covariates were controlled simultaneously within the correlation computation; thus, the reported correlation coefficients represent partial correlations. No prior revisualization of FDG or biomarker values was performed. To improve normality, WMH/TIV%, GFAP, and NfL values were log10-transformed prior to analysis. All analyses were conducted separately in MCSA and ADNI using all available data for each biomarker; sample sizes varied by biomarker and are reported in the figure legends.

Structural equation modeling

SEM analysis was conducted to explore the direct and indirect pathways linking WM metabolic signatures and its potential modifiers in both cohorts, using the R package lavaan. This framework combines multiple regression equations into a unified model, enabling the simultaneous estimation of direct, indirect, and total effects. Model variables were selected to maximize comparability between cohorts while retaining the largest possible sample size. These included demographic characteristics (age, sex, APOE-ε4 status, and vascular risk), brain pathological biomarkers (amyloid PET and WMH), WM metabolic signatures (EWM and AWM SUVR), and global cognition. Standardized path coefficients were estimated, and statistical significance was assessed using robust t-tests. Model fit was evaluated by χ² test, comparative fit index (CFI), Tucker–Lewis index (TLI), root mean square error of approximation (RMSEA), and standardized root mean square residual (SRMR). Good model fit was defined as non-significant χ², CFI/TLI > 0.95, and RMSEA/SRMR < 0.05. In addition, we computed the direct, indirect, and total effects of all predictors on EWM and AWM metabolism. This allowed us to compare the relative contributions of age, sex, APOE-ε4, vascular risk, amyloid, and WMH to each WM metabolic signature, and to highlight both consistent and divergent patterns across the MCSA and ADNI cohorts.

Microstructural basis underlying WM metabolism

To clarify whether there is a biological basis for the EWM and AWM metabolic signatures, we utilized available NODDI data in MCSA-CU participants (n = 235) to extract regional NDI values for the major WM fibers within two predefined regions of interest, with higher NDI values indicating greater microstructural integrity. Associations between regional WM NDIs and FDG-PET SUVR were assessed using partial correlation analyses, with the latter adjusted for age, sex, and APOE-ε4 status.

Prognostic value of WM metabolism in predicting cognitive decline beyond established imaging biomarkers

We investigated whether FDG-PET uptake in the WM metabolic signatures predicted longitudinal global cognitive change in both the MCSA and ADNI cohorts, while accounting for amyloidosis and vascular pathology. We fit a linear mixed effect model with global cognition as the dependent variable. The model included baseline age, follow-up time, sex, years of education, APOE-ε4 status, WM metabolism (both EWM and AWM regions), amyloid deposition, WMH burden, and the cycle number of the visit as fixed effects. Additionally, all main effects and their interactions with time were incorporated. Random effects for intercept and slopes were also included to account for initial differences and varying trajectories among participants.

Robustness analysis

To evaluate the robustness of the voxel-wise WM metabolic signatures, several processing- and modeling-level variations were performed. First, to assess the effect of image processing methods and potential GM spillover, voxel-wise regressions were repeated using PVC-uncorrected images and varying Gaussian smoothing kernel size (4, 6, and 8 mm). Second, to examine the influence of covariate adjustment, analyses were repeated with and without inclusion of age, sex, and APOE-ε4 status. Third, to address potential confounding by brain atrophy, total tissue volume to intracranial volume ratio (TTV/TIV) was included as an additional covariate, and analyses were repeated after excluding individuals with severe atrophy (TTV/TIV <mean–1.5 SD within age- and sex-stratified MCSA-CU distributions). In addition, cross-sectional regressions and longitudinal mixed-effects models were repeated using the MMSE score as the cognitive outcome.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

W.Z. and P.V. conceived and designed the study. W.Z. performed the data analysis. W.Z., S.R., J.T., and P.V. contributed to the interpretation of the results. S.A.P., H.J.W., and A.J.F. contributed to statistical methodology. M.L.S., C.G.S., and R.I.R. implemented the neuroimaging preprocessing pipelines and performed data and image quality control. W.Z. drafted the manuscript. M.M.M., R.C.P., J.G.-R., C.R.J., and V.J.L. provided critical feedback during manuscript revision. All authors approved the final manuscript and agreed to be accountable for all aspects of the work.

Peer review

Peer review information

Nature Communications thanks Julie Ottoy and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The MRI, PET, and associated clinical data from the MCSA are available under restricted access due to institutional review board (IRB) requirements and participant consent limitations. Access is limited to qualified academic and industry researchers for research purposes. Requests for access must be submitted to the MCSA Executive Committee via the Global Alzheimer’s Association Interactive Network (GAAIN) portal (https://www.gaaindata.org/partner/MCSA), where data use agreements and IRB documentation are reviewed prior to approval. Approved users are granted access under a formal data use agreement. The imaging and demographic data used in this study from the ADNI used in this study are publicly available in the ADNI database (https://adni.loni.usc.edu/) to qualified researchers upon registration. Source data generated in this study are provided with this paper. Source data are provided with this paper.

Code availability

All statistical analyses were conducted using R (version 4.4.0). No custom algorithms were developed. Image preprocessing was performed using an automated in-house pipeline built on SPM12. The analysis scripts used for image preprocessing are available from the corresponding author for research purposes.

Competing interests

J.G.-R. serves as a site principal investigator for clinical trials sponsored by Eisai and Cognition Therapeutics. V.J.L. serves as a consultant for Bayer Schering Pharma, Piramal Life Sciences, Life Molecular Imaging, Eisai Inc., AVID Radiopharmaceuticals, Eli Lilly and Company, PeerView Institute for Medical Education, and Merck Research, and receives research support from GE Healthcare, Siemens Molecular Imaging, and AVID Radiopharmaceuticals. R.C.P. serves as a consultant for Roche, Genentech, Eli Lilly and Company, Eisai Inc., Novartis, and Novo Nordisk; receives royalties from UpToDate; and receives honoraria for educational materials from Medscape. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-70707-6.