Correlations between white matter hyperintensities, magnetic resonance imaging findings, and cognition in the oldest old: a multimodal cross-sectional study

Abstract

Small vessel disease and brain atrophy are linked to cognition in later life, but whether these associations persist in the oldest old is unclear, as lesion burden and survivorship may change effect sizes. We analyzed 249 adults aged 89–94 years from the Kawasaki Aging and Wellbeing Project. White matter hyperintensities (WMHs) were segmented on fluid-attenuated inversion recovery; gray and white matter volumes were derived from T1-weighted magnetic resonance imaging; and peak width of skeletonized mean diffusivity (PSMD) indexed diffuse white matter damage. Global cognition was assessed with the Mini-Mental State Examination (MMSE). Robust multivariable regression related MMSE to age, education, sex, apolipoprotein E ε4 status, WMH burden, volumetric measures, PSMD, and a PSMD × education interaction. Voxel-wise lesion–symptom mapping and tract-based spatial statistics evaluated tract-level effects. Older age and higher WMH burden were associated with lower MMSE, whereas higher education was associated with higher MMSE. PSMD showed no independent main effect after covariate adjustment. An exploratory PSMD × education interaction was nominally significant in the primary model but attenuated in sensitivity analyses. Exploratory spatial analyses suggested that WMHs in the inferior fronto-occipital fasciculus (IFOF) and inferior longitudinal fasciculus, as well as higher axial diffusivity in the left IFOF, were associated with lower MMSE. Overall, the principal contribution of this study is confirmatory evidence, in a rare oldest-old cohort, that WMH burden and education remain associated with global cognition; spatial and interaction findings are hypothesis-generating and should not be interpreted as confirmatory evidence of tract-specific mechanisms.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1007/s11357-026-02295-8.

Introduction

With the increase in global life expectancy, more individuals are attaining advanced ages, leading to a growing societal burden of late-life cognitive decline [1]. Understanding the brain changes underlying cognitive aging is important for developing strategies to maintain cognitive health in the oldest old. Magnetic resonance imaging (MRI) has been instrumental in detecting structural brain correlates of age-related cognitive impairment [2]. In particular, the prevalence and volume of white matter hyperintensities (WMHs) on fluid-attenuated inversion recovery (FLAIR) MRI—which are radiological markers of cerebral small vessel disease—increase with age. Moreover, they are strongly associated with cognitive impairment and dementia risk [2]. A recent meta-analysis has confirmed that older adults with a greater WMH burden are at significantly higher risk of mild cognitive impairment (MCI) and dementia. Further, they experience a faster cognitive decline over time [2]. These findings underscore the role of WMH as a key contributor to cognitive aging. However, previous cohorts have commonly included the younger-old (e.g., individuals in their 60s–70s). Whether the same WMH–cognition association is applicable in oldest-old populations (aged ≥ 90 years) who often present with extensive brain pathologies yet exhibit substantial variability in cognitive outcomes remains uncertain. This knowledge gap in the geroscience literature motivated the current study.

Along with overt lesions such as WMHs, microstructural white matter integrity is an important substrate of cognitive aging. Diffusion tensor imaging (DTI) studies have shown that age-related degradation of white matter integrity (e.g., myelin loss) disrupts communication across neural networks and contributes to slower processing speed and executive dysfunction [3]. DTI metrics such as fractional anisotropy and mean diffusivity, which reflect axonal fiber integrity, are strongly correlated with cognitive performance in older adults [3]. Recently, the peak width of skeletonized mean diffusivity (PSMD)—an automated DTI-derived marker of diffuse white matter damage—has emerged as a sensitive indicator of cerebral small vessel disease. A higher PSMD is correlated with advanced age and reduced cognitive processing speed in aging [4]. According to some studies, PSMD is a predictor of cognitive performance. Further, it is considered better than conventional MRI markers of small vessel disease [5]. Incorporating such advanced structural MRI metrics (e.g., PSMD) may improve our ability to detect brain changes that affect cognition in the oldest old.

In addition to neuroimaging markers, individual risk and resilience factors influence cognitive trajectories in late life. The APOE ε4 allele—a well-known genetic risk factor for Alzheimer’s disease—is associated with greater white matter degeneration and steeper cognitive decline in older adults [6]. In contrast, a high amount of cognitive enrichment (educational, complex occupational, and social activities) is believed to build cognitive reserves, enabling some older adults to tolerate substantial neuropathology without clinical cognitive impairment [1]. Indeed, aging outcomes are heterogeneous. While some nonagenarians (aged 90–99 years) exhibit significant cognitive decline, others remain cognitively resilient or even outperform those decades younger [1]. The neural substrates underlying this resilience in extreme old age are not completely understood, as ultralong-lived individuals have been underrepresented in previous neuroimaging studies.

To address these gaps, the current research aimed to conduct a comprehensive cross-sectional MRI assessment of cognitive function in an oldest-old cohort (n = 249; mean age at Mini-Mental State Examination (MMSE) assessment, 91 years). We quantified WMH burden on FLAIR MRI and multiple structural MRI markers, including gray matter volume, normal-appearing white matter volume, and DTI-based PSMD, to determine their associations with global cognition (as assessed using the MMSE). Voxel-based lesion–symptom mapping (VLSM) and tract-based spatial statistics (TBSS) were also used to identify localized white matter lesions and microstructural changes associated with cognitive performance. This integrated approach allows us to test whether associations reported in younger-old samples—particularly those linking WMH burden to global cognition—remain detectable in a rare oldest-old cohort. In this context, our primary contribution is confirmatory rather than mechanistic. Analyses of tract specificity and PSMD-based moderation were included as exploratory components intended to refine future hypotheses for longitudinal work.

Rationale for focusing on the oldest old

Beyond age 90, lesion accumulation, multimorbidity, and survivorship can modify MRI–cognition associations and tract vulnerability. We therefore treat this work as confirmatory in an underrepresented age band, with exploratory analyses of tract specificity and reserve moderation to guide longitudinal follow-up.

Methods

Study participants and setting

This cross-sectional study utilized data from the Kawasaki Aging and Wellbeing Project in Kawasaki City, Japan [7]. In total, 249 oldest-old participants (mean age at the face-to-face MMSE assessment, 91.0 ± 1.3 years) were included in the analysis. MRI was performed subsequently. Among them, 118 were men and 131 were women. All participants were community-dwelling older adults who provided written informed consent. The study was conducted in accordance with the 1964 Declaration of Helsinki and its later revisions, and the study protocol was approved by the Ethics Committee of Keio University School of Medicine, Japan (approval ID: 20160297). Data on vascular comorbidities (hypertension, diabetes, dyslipidemia, smoking, and atrial fibrillation) were not collected and were therefore not analyzed.

Cognitive outcome assessment

Cognitive function was assessed using the MMSE, which was administered by trained clinical psychologists in private rooms. Although the MMSE lacks domain specificity (e.g., limited sensitivity to executive function, processing speed, or frontal-subcortical dysfunction) [8, 9], it was selected for this study due to its feasibility in extremely old participants—brief administration time, low participant burden, and widespread use in aging cohorts. The MMSE has also been robustly validated as a global cognitive screening tool in older populations (including nonagenarians) owing to its reliability and association with both clinical outcomes and neuroimaging markers [10, 11]. No clinical cutoff (e.g., dementia vs nondementia) was applied. Instead, the MMSE was treated as a continuous measure to maximize sensitivity across the full range of cognitive function. MMSE and MRI were obtained within 51 days of each other (median 6 days, range 2–51 days). This interval is reported to clarify the temporal proximity of the cognitive and imaging assessments in the present cross-sectional analysis.

MRI acquisition

MRI data were acquired at four sites (Y, D, K, and M) using Siemens 3 T scanners. Each protocol included axial FLAIR, DTI (b = 0 and 1000 s/mm², 12 directions), and high-resolution three-dimensional T1-weighted magnetization prepared rapid gradient echo. Sequence parameters were comparable across sites, with voxel sizes of approximately 0.7 mm in-plane for FLAIR, 1.7 × 1.7 × 3.0 mm for DTI, and approximately 0.9 × 0.9 × 1.0 mm for T1.

Quality control and harmonization

To address four-site acquisition, we implemented (i) standardized visual quality control (QC) with exclusion criteria for motion, susceptibility, and segmentation artifacts; (ii) quantitative motion and signal-to-noise ratio (SNR) metrics for diffusion; and (iii) harmonization using neuroComBat, an empirical Bayes method that removes scanner-related additive and multiplicative effects while preserving associations with biological covariates [12, 13]. We report site-wise distributions before and after harmonization and the variance explained by site (R²_site) pre- and post-harmonization. Datasets exceeding motion or artifact thresholds were excluded. A sensitivity model additionally included scanner site as a random intercept.

Image processing and quantification

WMH segmentation

WMHs were segmented from FLAIR and T1-weighted images using TrUE-Net, a deep learning model for WMH segmentation that has demonstrated high overlap with semi-manual reference labels and low false-positive rates in both development and independent validation studies [14, 15]. Probability maps were thresholded to generate binary lesion masks, and the total WMH volume was calculated by summing lesion voxels. For the analysis, WMH volumes were stratified into quartiles (Q1–Q4).

Lesion segmentation and manual QC

WMH were segmented from FLAIR/T1 using TrUE-Net with a probability threshold of p = 0.01, followed by morphological cleanup and a periventricular attrition correction to mitigate overestimation near ventricles. A single rater manually delineated n = 15 cases; automated–manual agreement showed a mean Dice overlap of 0.71 (range 0.62–0.81). Lesion masks were visually verified for continuity and alignment in Montreal Neurological Institute (MNI) space.

Brain volumetry

Structural MRI was processed with FreeSurfer (v.7.4.1) to extract the total cortical gray matter and cerebral white matter volumes. Lesional voxels were excluded from white matter volume estimates. FreeSurfer recon‑all failures 10, repaired 3. These measures served as indicators of atrophy and structural integrity in the regression models. Total and tissue volumes were normalized by intracranial volume (ICV) in sensitivity analyses. Excluding lesional voxels from white matter volume minimizes contamination by WMH; we clarify that reported white matter volume reflects normal-appearing white matter after lesion masking.

Diffusion MRI processing and PSMD calculation

Preprocessing

Diffusion MRI data (b = 1000 s/mm²) were denoised, Gibbs-corrected (MRtrix3), distortion-corrected (Synb0-DISCO), and registered to the MNI space using Advanced Normalization Tool. DTI was acquired with 12 directions (b = 1000 s/mm²), which we acknowledge as a limitation for tensor precision; preprocessing mitigated noise and distortions. To address multisite acquisition, diffusion metrics were harmonized using neuroComBat, which has been validated for structural and diffusion MRI and shown to reduce site effects while preserving biologically meaningful associations [12, 13]. Motion was quantified using the median volume-to-volume relative root-mean-square (RMS) displacement (mm) from FSL eddy’s eddy_movement_rms file (column 2, relative to the previous volume). SNR was computed from Marchenko–Pastur principal component analysis noise maps (MRtrix3 dwidenoise with -noise), defined as SNR = mean b = 0 signal within a white matter mask divided by the estimated noise standard deviation σ. Subjects with absolute RMS ≥ 2.0 mm were excluded or flagged. Site explained R²_site = 0.09 of diffusion-metric variance pre-harmonization and 0.01 post-harmonization; associations with age, sex, education, and APOE ε4 status were preserved. Diffusion head-motion median = 0.043 mm (interquartile range 0.009–0.175); SNR median = 36.4; four datasets were excluded for motion ≥ 2.0 mm or artifacts.

DTI metrics and skeletonization

Diffusion tensors were fit with FMRIB’s software library to generate fractional anisotropy (FA), MD, axial diffusivity (AD), and RD maps. Using TBSS, the FA images of the participants were aligned to the FMRIB FA template and projected onto a mean white matter skeleton, ensuring voxel-wise correspondence across tracts. The skeleton mask followed TBSS conventions with an FA threshold of 0.2.

PSMD

PSMD was computed on the white matter skeleton as the 95th minus the 5th percentile of MD (P95−P5). A larger PSMD reflected a greater heterogeneity in diffusivity and indicated more diffuse white matter damage. Individual PSMD values were z-standardized for use in the regression analyses.

Voxel-wise analyses

TBSS

Skeletonized DTI data were analyzed with voxel-wise GLMs to examine the associations between MMSE scores and diffusion metrics (FA, MD, AD, and RD), after controlling for age, sex, education, and APOE4. Significance was assessed using 5000 permutations with TFCE, family-wise error-corrected at p < 0.05. This process identified white matter tracts with microstructural integrity related to cognition in extremely elderly individuals.

VLSM

Binary WMH lesion masks were normalized to the MNI space and compared voxel-wise for differences in MMSE scores between the lesion-present and lesion-absent groups, after adjusting for the same covariates. Voxels with insufficient lesion counts were excluded from the analysis. The threshold for significance was p < 0.001 (uncorrected), highlighting regions where focal lesions predicted poorer cognition.

Taken together, TBSS and VLSM provided complementary evidence on whether global lesion burden or specific lesion locations drive cognitive impairment in the oldest old.

Statistical analysis

A multivariable robust linear regression analysis was performed to examine the predictors of MMSE. Independent variables included WMH volume, gray matter volume, white matter volume, PSMD, years of education (proxy for cognitive reserve), age, sex, and APOE ε4 status (binary). APOE ε4 status was coded as 1 for carriers (≥ 1 ε4 allele) and 0 for noncarriers. An interaction term (PSMD × education) tested whether microstructural damage moderated reserve effects.

WMH and volumetric measures were stratified into quartiles. Both ordinal (1–4) and categorical (Q1 reference) models were tested to assess linear versus threshold effects. Robust regression with Huber weighting (Python Statsmodels) reduced outlier influence. The predictors met linearity assumptions, multicollinearity was low (VIF < 2.0), and residuals approximated normality.

Both model versions yielded consistent results, supporting a monotonic association between a higher WMH burden and lower cognition. We report unstandardized coefficients with SEs, two-tailed α = 0.05, and model fit indices (R² and adjusted R²). In addition, to aid interpretation of effect magnitude, we report predictor-level explained variance for the primary model as partial R² or an equivalent variance-partitioning metric appropriate to the modeling framework. To probe nonlinearity, WMH burden was modeled with restricted cubic splines (k = 5 knots) in addition to ordinal/categorical encodings. Sensitivity analyses included (i) models with site as a random intercept, (ii) models including ICV, and (iii) exclusion of high-motion diffusion datasets.

Results

Overall cognitive performance and lesion burden

Across the 249 community-dwelling adults aged 89–94 years, the mean MMSE score was 26.0 ± 3.1, and 22.5% of participants had scores ≤ 23. We did not apply any MMSE cutoff for inclusion. Total WMH volume varied widely; participants were divided into quartiles (Q1–Q4) for analysis. Demographic, clinical, and imaging characteristics are summarized in Table 1. Vascular comorbidity variables were unavailable in this cohort and are not reported.

Table 1.

Characteristics of the participants

Characteristics of the participants	Mean (SD) or n (%)
Total number of participants	249
Age at face-to-face assessment (years)	91 (1.3)
Sex (male/female)	118/131 (47.4%/52.6%)
MMSE score	26.0 (3.1)
MMSE score ≤ 23 (dementia group)	56 (22.5%)
Years of education	12 (3.3)
APOE ε4 carrier	24 (9.6%)
WMH volume (mm3)	30,105 (23,320)
PSMD	0.00066 (0.00052)

APOE ε4 apolipoprotein E epsilon 4 allele, MMSE Mini-Mental State Examination, PSMD peak width of skeletonized mean diffusivity, SD standard deviation, WMHs white matter hyperintensities

In multivariable robust regression models, age, WMH burden, and education emerged as the primary correlates of global cognition. In the ordinal WMH model, older age (B =  − 0.52, p < 0.001) and a greater WMH burden (B =  − 0.40, p = 0.021) were associated with lower MMSE scores, whereas higher educational level (z-score) was associated with better cognition (B =  + 0.60, p = 0.002). Sex, APOE ε4 carrier status, and quartile-based total gray and white matter volumes were not significantly related to MMSE scores. PSMD showed no independent main effect after adjustment for these covariates. Predictor-level explained variance indicated that age, education, and WMH burden accounted for the largest share of explained variance in MMSE, whereas PSMD, tissue-volume quartiles, sex, and APOE ε4 contributed comparatively little additional explanatory value (Table 2 and Table S1).

Table 2.

Robust regression results of potential predictors of MMSE scores in our cohort (with unstandardized B and p values)

Predictors	Ordinal WMH model			Categorical WMH model
	B, p value	95% CI (lower, upper)	Partial R2 (approx)	B, p value	95% CI (lower, upper)	Partial R2 (approx)
Age	− 0.52, < 0.001	− 0.82, − 0.23	0.050	− 0.57, < 0.001	− 0.86, − 0.27	0.056
Educational level (z-score)	+ 0.60, 0.002	0.22, 0.98	0.039	+ 0.54, 0.007	0.15, 0.93	0.030
Sex (male = 1)	+ 0.04, 0.920	− 0.83, 0.93	0.000	+ 0.01, 0.987	− 0.89, 0.90	0.000
APOE ε4 carrier	+ 0.95, 0.121	− 0.25, 2.14	0.010	+ 0.96, 0.121	− 0.25, 2.17	0.010
WMH quartile (ordinal)	− 0.40, 0.021	− 0.75, − 0.06	0.022	—	—	—
WMH Q2 vs Q1	—	—	—	+ 0.35, 0.504	− 0.67, 1.37	0.002
WMH Q3 vs Q1	—	—	—	− 0.93, 0.081	− 1.97, 0.11	0.013
WMH Q4 vs Q1	—	—	—	− 0.96, 0.085	− 2.05, 0.13	0.013
White matter volume (quartile)	+ 0.12, 0.592	− 0.32, 0.56	0.001	+ 0.07, 0.753	− 0.38, 0.52	0.000
Gray matter volume (quartile)	− 0.05, 0.815	− 0.49, 0.39	0.000	− 0.01, 0.964	− 0.45, 0.43	0.000
PSMD (z-score)	− 0.08, 0.661	− 0.44, 0.28	0.001	− 0.09, 0.635	− 0.46, 0.28	0.001
PSMD × education	− 0.39, 0.046	− 0.78, − 0.01	0.017	− 0.39, 0.049	− 0.79, 0.00	0.017

Coefficients are unstandardized; education and PSMD were z-scored prior to analysis. Ordinal WMH model includes WMH as an ordinal quartile variable (per 1-quartile increase), whereas categorical WMH model includes dummy indicators for WMH quartiles (reference = Q1). APOE ε4 coding: 1 = carrier, 0 = noncarrier. A positive coefficient should not be interpreted as protective in the absence of statistical significance

APOE ε4 apolipoprotein E epsilon 4 allele, approx approximate, B unstandardized regression coefficient, CI confidence interval, MMSE Mini-Mental State Examination, p p value, PSMD peak width of skeletonized mean diffusivity, Q1, Q2, Q3, Q4 quartiles 1, 2, 3, and 4, R² coefficient of determination, WMHs white matter hyperintensities, z-score standardized score

When WMH was modeled categorically with Q1 as the reference, the overall pattern was similar. Age (B =  − 0.57, p < 0.001) and education (B =  + 0.54, p = 0.007) remained significant predictors, whereas sex and APOE ε4 status continued to show null associations. Higher WMH quartiles (Q3 and Q4 vs Q1) tended to be associated with lower MMSE scores (both p = 0.08), but these contrasts did not reach the conventional significance threshold. Volumetric gray and white matter measures and PSMD main effects again remained nonsignificant. The full set of coefficients from the ordinal and categorical WMH models is summarized in Table 2.

Exploratory PSMD × education interaction

Although PSMD did not show a robust independent association with MMSE, we observed a nominal PSMD × education interaction in the primary model (B =  − 0.39; 95% confidence interval (CI) − 0.78 to − 0.01; p = 0.046). This interaction suggests that the positive association between education and MMSE may be smaller at higher PSMD values. However, in sensitivity models, the interaction estimate was similar in magnitude but borderline (B =  − 0.37; 95% CI − 0.76 to 0.02; p = 0.062), and we therefore interpret this moderation pattern as exploratory and hypothesis-generating rather than definitive.

To explore potential nonlinearity in the WMH–cognition relation, we compared an ordinal model with a model using restricted cubic splines (five knots). The likelihood-ratio test suggested some deviation from strict linearity (χ²(3) = 8.06, p = 0.045), consistent with a graded but possibly threshold-like relationship between WMH burden and MMSE.

Sensitivity analyses

Model fit was modest but consistent across specifications (R² = 0.176; adjusted R² = 0.140; n = 243). Fig. S1 shows the distributions of WMH and PSMD, and Fig. S2 shows site-wise distributions before and after harmonization. The pattern of associations was largely unchanged when we added scanner site as a random intercept and intracranial volume as an additional covariate. In these sensitivity models, the WMH effect remained negative but weaker and nonsignificant [B = –0.22, 95% CI − 0.59 to 0.15, p = 0.241], whereas the education effect remained robust [B = 0.65, 95% CI 0.25 to 1.05, p = 0.002]. The PSMD × education interaction was of similar magnitude but only borderline significant [B =  − 0.37, 95% CI − 0.76 to 0.02, p = 0.062]. Simple-slope (Johnson–Neyman) plots illustrating the moderation pattern are provided in Fig. S3 (exploratory).

Exploratory voxel-wise lesion and tract-level analyses

VLSM and TBSS provided complementary tract-level information. At an uncorrected threshold of p < 0.001, VLSM revealed clusters in the bilateral inferior fronto-occipital fasciculi (IFOF) and the left inferior longitudinal fasciculus (ILF), where the presence of WMH was associated with lower MMSE scores (Fig. 1A). To facilitate interpretation of the VLSM findings, Fig. 1B plots MMSE distributions by lesion status within these clusters, illustrating that participants with WMH in the IFOF/ILF tend to have lower MMSE scores than those without lesions. These findings should be regarded as exploratory because they were not corrected for multiple comparisons.

Fig. 1.

Exploratory voxel-wise lesion–symptom mapping (VLSM) of white matter hyperintensities (WMH) versus Mini-Mental State Examination (MMSE). A Threshold: uncorrected p < 0.001; interpret as hypothesis-generating spatial findings rather than confirmatory evidence of tract-specific effects. Color bar: t values 3.12–5.15. Statistical maps are overlaid on the Montreal Neurological Institute 152 (MNI152) template (McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University; open-access). Blue denotes negative associations (greater WMH burden associated with lower MMSE). B Boxplots of MMSE scores stratified by lesion status (lesion-present vs lesion-absent) within the main VLSM clusters in the bilateral inferior fronto-occipital fasciculi (IFOF) and left inferior longitudinal fasciculus (ILF). Boxes indicate the interquartile range with the median line; whiskers represent 1.5 × interquartile range, and dots depict individual participants. Lesion-present groups show lower MMSE scores than lesion-absent groups, illustrating the direction and magnitude of the VLSM effects

In the TBSS analysis of diffusion metrics, a focal cluster in the left IFOF showed a significant association between AD and MMSE after family-wise error correction (p < 0.05), such that higher AD was associated with lower MMSE scores (Fig. 2A). Figure 2B provides a complementary quantitative view by plotting cluster-averaged AD against MMSE, showing that higher AD values within the left IFOF were associated with lower MMSE scores across individuals. No FWE-corrected effects were found for fractional anisotropy, mean diffusivity, or radial diffusivity. Accordingly, this TBSS result should be interpreted as a limited exploratory spatial finding rather than as confirmatory evidence of a specific tract-level mechanism. Color bar scales, anatomical overlays, and threshold details are provided in the figure legends.

Fig. 2.

Exploratory tract-based spatial statistics (TBSS) results. A A focal cluster in the left inferior fronto-occipital fasciculus (IFOF) showed a negative association between axial diffusivity (AD) and Mini-Mental State Examination (MMSE) score (threshold-free cluster enhancement [TFCE], family-wise error [FWE]–corrected p < 0.05). This finding should be interpreted as a limited exploratory spatial result and not as confirmatory evidence of a specific tract-level mechanism. Results are shown as corrected 1 − p (“corrp”) values thickened with tbss_fill and overlaid on the mean fractional anisotropy (FA) skeleton (green), representing the centers of major white matter tracts. A yellow anatomical overlay from the Johns Hopkins University (JHU) white matter tractography atlas labels the left inferior fronto-occipital fasciculi (IFOF) for anatomical localization (Montreal Neurological Institute 152 [MNI152] background; coronal, sagittal, and axial views). B Scatterplot of cluster-averaged AD values within the significant left IFOF cluster against MMSE scores. Each dot represents one participant; the line depicts the fitted linear association (adjusted for age, sex, education, and apolipoprotein E ε4 status), with a negative slope indicating that higher AD is associated with lower global cognition. L, left

Primary (confirmatory) analyses indicated that MMSE was most consistently related to age, overall WMH burden, and education. Exploratory spatial analyses provided hypothesis-generating tract-level signals (uncorrected VLSM, a limited focal TBSS effect) that may inform future work on long association fibers (e.g., IFOF/ILF), but these findings should not be interpreted as confirmatory evidence of tract-specific or network-level mechanisms and require replication in independent samples.

Discussion

In this community-based cohort of adults aged 89–94 years, the main confirmatory finding was that greater WMH burden and older age were associated with lower MMSE scores, whereas higher education was associated with better MMSE performance. Secondary analyses yielded an exploratory PSMD × education interaction and tentative tract-level signals in long association fibers. These secondary findings are hypothesis-generating and should not be interpreted as mechanistic evidence. Below, we discuss these themes in turn and consider their implications within the broader geroscience literature.

Vascular lesion burden and cognition in the oldest old

Our findings confirm that WMH burden remains an important correlate of global cognition in the tenth decade of life. Across regression models, higher WMH burden was associated with lower MMSE scores, with some evidence for a nonlinear pattern, in which cognitive performance declines more steeply once lesion burden exceeds lower quartiles. The effect size of WMH (unstandardized B ≈ − 0.4 per quartile in the ordinal model) is comparable to that reported in meta-analyses and large cohorts enriched for younger-old adults in their 60 s and 70 s, in which WMH burden predicts both cross-sectional cognitive differences and subsequent decline [2, 12]. The similarity in effect sizes suggests that small vessel disease continues to explain a meaningful portion of cognitive variance even among nonagenarians.

This result supports the view that WMHs are not merely incidental markers of aging but represent clinically important small vessel injury that remains linked to stroke, dementia, gait disturbance, and mortality throughout late life [2, 16]. In our sample, substantial heterogeneity in WMH burden was observed despite the narrow age range, and higher burden was associated with worse global cognition. At the same time, the overall variance explained by our models was modest, indicating that WMHs and age capture only part of the multifactorial processes driving cognitive trajectories in the oldest old. Other vascular, neurodegenerative, systemic, and psychosocial influences likely contribute to the residual variance in cognitive outcomes.

Cognitive reserve, PSMD, and the limits of compensation

The second major pattern concerned education and microstructural white matter damage. Consistent with cognitive reserve theory, higher educational attainment was associated with better MMSE performance. The magnitude of this association (B ≈ + 0.6 per SD of education) is in line with, or somewhat larger than, estimates from mixed-age cohorts and suggests that differences in lifetime cognitive enrichment continue to matter even in advanced old age [1, 17–23]. Nonagenarians with higher education in our sample tended to perform better on a global cognitive screening measure despite the presence of extensive age-related pathology.

PSMD, a diffusion-derived index of diffuse white matter damage, did not show a robust independent main effect on MMSE after adjustment for other covariates. However, it moderated the association between education and cognition, such that the cognitive advantage linked to higher education was smaller at higher PSMD values. This moderation pattern (nominal in the primary models and borderline in sensitivity analyses) is consistent with the possibility that diffuse white matter damage may reduce the observable association between education and global cognitive screening performance. Given that the MMSE is not optimized for executive function or processing speed, we avoid strong mechanistic interpretations (e.g., compensation “limits”) and view this interaction as exploratory, motivating future longitudinal and domain-specific cognitive testing. The finding also fits with prior work indicating that PSMD is particularly sensitive to processing speed and executive function, domains that are only partially captured by the MMSE [4, 5, 24].

Several methodological and cohort-specific features may explain why PSMD itself did not show a strong main effect. The DTI protocol used 12 directions at b = 1000 s/mm², which restricts tensor precision. The age range was narrow, and participants survived to very advanced ages and were healthy enough to complete MRI, which may have reduced variability in both PSMD and cognitive scores. Furthermore, we relied on the MMSE as a single measure of global cognition; more detailed neuropsychological testing might reveal clearer PSMD–behavior relations, particularly in speeded and executive domains. Given these constraints and the modest effect size, we view the PSMD × education interaction as hypothesis-generating evidence that diffuse white matter damage can erode the benefits of reserve rather than as definitive proof of a “ceiling” on compensation.

Tract-level signals and network vulnerability

Voxel-wise analyses provided additional, spatially resolved context for the global associations. VLSM maps highlighted clusters in the bilateral IFOF and left ILF, where WMH presence was associated with lower MMSE scores at an uncorrected threshold. TBSS analyses identified a focal FWE-corrected cluster in the left IFOF, in which higher AD was related to lower MMSE scores. Although these long association pathways have been implicated in integrative cognitive functions, the present spatial analyses were exploratory and hypothesis-generating and were not designed to establish tract-specific or network-level mechanisms. Accordingly, the observed localization should be interpreted cautiously as spatial context for the global findings rather than as mechanistic evidence [3, 4, 25, 26].

Relative to prior work that has focused primarily on global WMH volume or whole-brain DTI averages, our spatial findings add topographical detail that may help prioritize regions of interest for future study. At the same time, several caveats are essential. The VLSM clusters were not corrected for multiple comparisons and should be treated as exploratory. Only a single small cluster survived FWE correction in TBSS, and we did not examine specific cognitive domains that might map more tightly onto particular tracts. Thus, the spatial findings are best viewed as hypothesis-generating observations to be tested in larger samples with more comprehensive cognitive batteries.

APOE ε4 in very late life

In this cohort, APOE ε4 carrier status was not significantly associated with MMSE scores, and we did not perform dedicated analyses linking APOE ε4 to imaging markers such as WMH, PSMD, or tract-level diffusion metrics. Accordingly, the present data do not allow firm conclusions about ε4-related imaging pathways or about whether vascular lesion burden mediates ε4 effects on cognition in the oldest old. Existing studies in younger and mixed-age cohorts suggest that APOE ε4 is associated with greater WMH burden and faster WMH accumulation [6, 18–21], whereas work in centenarians indicates lower ε4 prevalence and a diminished cognitive impact of the allele [17, 22, 27, 28]. Our null findings regarding ε4 are compatible with an attenuated effect in very late life but must be interpreted cautiously given the relatively small number of carriers and the cross-sectional design. In this context, the APOE discussion should be understood as providing background rather than as being driven by strong genotype–phenotype associations in our data.

Methodological considerations and future directions

Several limitations qualify our findings and point to directions for future research. The cross-sectional design precludes determination of temporal ordering between MRI changes and cognitive decline; we cannot ascertain whether higher WMH burden or PSMD values preceded or resulted from cognitive deterioration. Longitudinal studies with repeated MRI and multidomain neuropsychological testing are needed to examine whether WMH, PSMD, and tract-level metrics predict subsequent rates of change in cognition.

Selection and survivorship biases are also likely. Participants were sufficiently healthy and mobile to undergo 3 T MRI at collaborating sites, which may underrepresent frailer nonagenarians and institutionalized elders. As a result, pathology–cognition associations in this cohort could underestimate the strength of relationships present in the broader population, and null findings (e.g., some diffusion metrics and APOE ε4) may reflect restricted ranges and limited power.

Cognition was assessed solely with the MMSE. Although practical for extremely old participants and predictive of important clinical outcomes, the MMSE is relatively insensitive to subtle deficits and to domain-specific impairments in executive function and processing speed [8–11, 29]. Future work incorporating comprehensive cognitive batteries would likely yield a more nuanced picture of how vascular lesions, microstructural damage, and reserve relate to specific cognitive domains.

Importantly, vascular risk factor data (e.g., hypertension, diabetes, and smoking) were unavailable, limiting our ability to adjust for potential confounding and to contextualize WMH burden in relation to systemic vascular exposures. Because such vascular factors are plausibly associated with both greater WMH burden and poorer cognition, their absence may have inflated the apparent direct association between WMH and MMSE. At the same time, treatment heterogeneity and survivor selection in this very old cohort could also attenuate observed associations. Accordingly, the present data do not support causal attribution of cognitive differences to WMH burden alone. We also lacked neurodegenerative biomarkers (e.g., amyloid/tau measures), precluding separation of purely vascular from mixed-pathology contributions that are common in very late life [25, 26, 30]. In addition, we did not systematically quantify MRI markers such as microbleeds or cortical microinfarcts. Diffusion-based indices like PSMD and AD are biologically nonspecific and should be interpreted as global indicators of structural injury rather than as markers of particular cellular mechanisms.

Although we implemented rigorous four-site quality control and neuroComBat harmonization, residual site effects and technical limitations cannot be ruled out. The 12-direction DTI protocol constrains angular resolution and tensor estimation; choices regarding PSMD computation, skeletonization, and QC thresholds may have influenced sensitivity. Replication with higher-direction diffusion protocols, harmonized acquisition schemes, and pre-registered analytic plans would strengthen confidence in the present results.

Finally, our cohort comprised urban Japanese elders. Educational systems, vascular risk exposures, diet, and lifestyle patterns differ across populations, and it is uncertain how generalizable these findings are to other ethnic and cultural groups [17, 22, 25, 30]. Cross-cohort and cross-national comparisons will be important for identifying which MRI–cognition relationships represent robust features of very late-life aging and which are context dependent.

We are currently pursuing longitudinal follow-up within this cohort, including repeat MRI and more detailed cognitive testing, to model individual trajectories of WMH and PSMD against cognitive change. These analyses will also incorporate additional reserve-related proxies (such as occupational complexity, social engagement, and lifestyle factors) to clarify how structural brain changes and life-course enrichment jointly shape cognitive outcomes at the limits of the human lifespan.

Clinical and public health implications

Despite these limitations, the overall pattern of results carries several practical implications. First, WMH burden and microstructural white matter damage—likely reflecting cumulative vascular risk—were closely linked to global cognition, whereas education-related reserve provided only partial compensation. This pattern aligns with evidence that controlling hypertension, diabetes, and other vascular risk factors, promoting physical activity, and avoiding smoking help reduce small vessel disease and support cognitive health [2, 16, 23, 31]. Our findings suggest that such vascular risk management remains relevant even for individuals who reach very advanced ages.

Second, the moderation of educational benefits by PSMD implies that high educational attainment does not fully offset the impact of extensive white matter damage. Clinically, this underscores the value of incorporating both FLAIR-based WMH assessments and, where feasible, diffusion-derived indices into the evaluation of very old adults, including those with high educational attainment. Identifying individuals with high lesion burden and reduced white matter integrity may help clinicians anticipate vulnerability to further decline and tailor monitoring and counseling accordingly.

Conclusion

In a narrow age band of community-dwelling adults aged 89–94 years, this study provides confirmatory evidence that WMH burden, age, and education remain associated with global cognition in the oldest old. The pattern and approximate effect sizes of these associations resemble those reported in younger-old cohorts, supporting the continued relevance of small vessel disease and life-course cognitive enrichment in very late life. Exploratory interaction and tract-level findings should be interpreted cautiously and require replication. These findings support a life-course emphasis on vascular risk management and cognitive enrichment, while motivating longitudinal multimodal studies to elucidate how vascular, neurodegenerative, and network-level processes jointly determine who reaches extreme old age with preserved cognitive function.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

M.M. and Y.A. initiated the research project. All authors contributed to the study’s conception and design. R.S., J.H., and T.S. prepared the materials and collected the data. R.U. conducted the data analysis and drafted the initial manuscript. Y.A. also contributed to manuscript preparation, secured funding, and supervised the overall project. All authors critically revised the manuscript and approved the final version.

Funding

This study was supported by JST SPRING (grant no. JPMJSP2123), Japan Society for the Promotion of Science (grant nos. 18H03055, 23H03337, 24K20224), JST Research Complex Program (grant no. JP15667051), AMED (grant nos. JP20jm0210051h0004, 23zf0127007h0002), Keio Global Research Institute, Kanagawa Institute of Industrial Science and Technology, the Univers Foundation (grant no. 21-01-194), and the Sasakawa Peace Foundation (grant no. 2022A-003).

Data availability

The datasets analyzed in this study contain sensitive information from vulnerable populations and are therefore subject to ethical and legal restrictions that prevent public deposition. Access to these datasets may be granted following a reasonable request to the principal investigator (Y. Arai), pending approval by the Research Ethics Committee for Clinical Research at Keio University School of Medicine.

Declarations

Ethics approval and consent to participate

The Ethics Committee of Keio University approved this study (approval number 20160297). All participants or their legally authorized representatives provided written informed consent for study participation. The current study was conducted in accordance with the Declaration of Helsinki. Data were anonymized for analysis and presentation.

Conflict of interest

Y. Arai received research support from Ezaki Glico Co., Ltd., Johnson & Johnson K.K., and Tuning Folk Bio Japan for this study. The other authors declare no competing interests.