Long-term prognosis and educational determinants of brain network decline in older adult individuals

The study investigates whether educational attainment is linked to longitudinal changes in large-scale functional brain network organization during adulthood and whether such changes predict impending cognitive decline. Education is a strong social determinant of health, with lower attainment associated with increased risk of mental health disorders and dementia in late life. Prior efforts to connect education and environment to brain structural change and pathological burden have produced mixed results. The authors hypothesize that an individual’s resting-state functional network organization—specifically system segregation reflecting the modularity and functional specialization of brain systems—indexes functional integrity and that its degradation with age may forecast cognitive impairment beyond neurodegeneration and AD pathology. The work aims to clarify how education relates to trajectories of brain network aging and to test the prognostic value of brain network change for future dementia severity.

The paper reviews established AD biomarkers within the A/T/N framework (amyloid, tau, neurodegeneration) and notes that individuals with similar biomarker profiles can show different clinical outcomes, implying unaccounted moderators. Incorporating brain function into aging and AD models has been challenging, but resting-state functional correlations (RSFC) provide a task-free window into large-scale network organization. RSFC networks exhibit modular architecture (system segregation), which supports functional specialization and varies across the lifespan, is associated with cognitive ability, and is altered in disease. Prior studies show age-related reductions in system segregation (dedifferentiation) and cross-sectional links between lower socioeconomic status and lower segregation in midlife. Some reports also find reduced segregation in AD compared with age-matched controls. This body of work motivates testing education-related differences in longitudinal changes of system segregation and their relationship to clinical progression.

Design and participants: Longitudinal cohort from the Knight Alzheimer Disease Research Center (Washington University in St. Louis). Inclusion required at least two resting-state fMRI sessions plus demographic and education data. After rigorous preprocessing and quality control, 265 adults (154 female), baseline age 45–86 years (mean 67.01, s.d. 9.26), were analyzed. Each participant had 2–5 MRI sessions over 0.88–9.24 years, with most sessions spaced at ~1-year intervals. Clinical, genetic, and biomarker data closest to baseline MRI were used when not collected concurrently. Education grouping: Self-reported years of education were dichotomized: college+ (≥16 years) versus below college (<16), reflecting the socioeconomic and health advantages associated with degree completion. Imaging acquisition and preprocessing: For each session, T1-weighted structural MRI supported FreeSurfer 5.3 cortical surface reconstruction and cortical thickness estimation with manual QC and edits. Resting-state fMRI (EPI; TR=2200 ms, TE=27 ms, 4 mm isotropic voxels; typically two runs of 164 volumes) was preprocessed using Nipype (FSL, SPM): slice-timing correction, motion correction, alignment to T1, intensity scaling, nuisance regression (global, white matter, ventricle signals and derivatives; Friston 24 motion regressors), motion scrubbing (FD>0.3 mm; interpolation), bandpass filtering (0.009–0.08 Hz), removal of interpolated frames, and global signal regression. Data were surface-mapped to fs_LR (32k), smoothed (σ=2.55), and sessions were standardized to 100 frames to control for varying data quantity. Network construction and measures: A 349-node cortical parcellation (surface-based, boundary-informed) was used, with nodes labeled by functional system membership (based on Power et al.). For each session, node time series were averaged within nodes; pairwise Pearson correlations were computed, Fisher z-transformed; negative correlations (potentially introduced by global signal regression) were set to zero. System segregation was computed as the proportional difference between mean within-system and between-system correlations across all systems. Structural measures included whole-brain mean cortical thickness and mean hippocampal volume (adjusted for intracranial volume). Clinical and biomarker measures: Dementia severity was assessed longitudinally via Clinical Dementia Rating Sum of Boxes (CDR-SB; range 0–18), with additional analyses using global CDR. AD-related pathology included dichotomous variables for elevated cortical amyloid (PiB PET SUVR>1.42) and elevated CSF phosphorylated tau 181 (>67 pg/mL). APOE ε4 carrier status was recorded. Health covariates included a composite cardiovascular risk factor (BMI>30; history of hypertension, hypercholesterolemia, cardiovascular incidents), Geriatric Depression Scale (GDS), and traumatic brain injury (TBI) history. Socioeconomic context measures included occupation-based socioeconomic index (SEI), neighborhood median household income, and Area Deprivation Index (ADI). Statistical analysis: Linear mixed-effects models (R lme4) modeled longitudinal changes. For brain outcomes, normalized time from baseline (within-participant), education group (between-participant), and age at baseline (between-participant) and their interactions predicted system segregation, controlling for sex and in-scanner motion; race and additional covariates were added in sensitivity models. Multiple comparisons for the two brain measures (segregation, cortical thickness) used a corrected significance threshold of P<0.025. For clinical outcomes, mixed models predicted longitudinal CDR-SB with interactions of time, age at baseline, and change in system segregation (difference last–baseline), controlling for sex, head motion (baseline and last), time between scans, and education group, and testing moderation by AD pathology, APOE status, and structural change. Block-level RSFC changes were assessed using system-level matrices (within- and between-system blocks) with paired t-tests (first vs last time point) and permutation-based significance (n=10,000) with FDR correction.

- Education differentiates socioeconomic context: College+ participants had higher occupation-based SEI, lived in higher-income neighborhoods, and had lower ADI (all P≤0.002), confirming the socioeconomic distinction of degree completion. - Longitudinal system segregation: Mixed models showed a main effect of education (F(2,269)=5.073, P=0.025) with higher segregation in college+ and a main effect of baseline age (F(1,269)=7.029, P=0.008; older age associated with lower segregation). Significant interactions included education×age (F(1,270)=4.949, P=0.027) and time×education (F(1,136)=5.756, P=0.017). Critically, a three-way interaction time×education×baseline age was significant (e.g., controlling sex and motion: F(1,193)=6.814, P=0.010), indicating that older adults (≥~69 years) without a college degree showed significant declines in system segregation over time, whereas slopes in college+ older adults were flatter and not reliably negative. Results were robust to adjusting for race and sex interactions. - Independence from AD risk, pathology, and health factors: The education–segregation relationship persisted after controlling for baseline clinical status (CDR), APOE ε4, AD pathology (amyloid and/or CSF pTau), cardiovascular health, GDS, TBI, and combinations thereof (Table 2; three-way interaction remained significant or marginally significant across models; e.g., with all covariates, P=0.073). - Independence from structural change: Cross-sectionally, segregation correlated weakly with mean cortical thickness after age control (partial r(262)=0.066, P=0.284). The three-way interaction predicting segregation remained significant after controlling for longitudinal cortical thickness (F(1,195)=5.915, P=0.016). By contrast, longitudinal cortical thickness decline showed age effects but no time×age×education interaction (F(1,178)=0.003, P=0.954), indicating structural thinning did not vary by education. - System-specific RSFC changes: Older adults in both education groups showed time-related RSFC changes, but ‘below college’ older adults had greater increases in between-system correlations involving default mode, frontal-parietal control, and memory-retrieval systems, and decreases in within-system correlations (e.g., cingulo-opercular and visual), contributing to reduced segregation. - Prognosis of cognitive decline: Decline in system segregation predicted future increases in CDR-SB beyond the last MRI, up to 10.1 years post-scan. A mixed model showed a significant three-way interaction time×baseline age×change in segregation (F(1,258)=4.292, P=0.039); with AD pathology and APOE added, change in segregation remained significant (F(1,242)=7.957, P=0.005), while AD pathology (F(1,233)=3.277, P=0.072) and APOE (F(1,24)=3.057, P=0.082) were marginal. Higher-order interactions including segregation with APOE/pathology were not significant (P>0.108). Effects persisted when adjusting for longitudinal changes in cortical thickness or hippocampal volume; replacing segregation change with structural changes did not predict CDR-SB (F(1,250)<0.785, P>0.377). Accounting for concurrent CDR-SB change during the MRI interval did not alter the predictive association (F(1,250)=7.273, P=0.007). - Education did not moderate the segregation–outcome link: The interaction education×time×age×segregation change predicting CDR-SB was not significant (F(1,250)=0.317, P=0.574), indicating that once network decline occurs, its clinical impact is similar regardless of educational attainment.

The study shows that educational attainment is related to how functional brain network organization changes with age: older adults without a college degree exhibit greater longitudinal declines in resting-state system segregation than college-educated peers. This dedifferentiation concentrates in association systems supporting integrative processing. Importantly, declining segregation forecasts worsening dementia severity beyond the imaging window and independent of APOE ε4, amyloid/tau pathology, and structural atrophy, suggesting that functional network reorganization captures aspects of brain health not reflected in traditional A/T/N biomarkers. While education relates to the trajectory of network aging, it does not buffer the clinical consequences once segregation declines, aligning with reserve concepts in which functional network organization may be a proximal brain measure of reserve. The results underscore the need to incorporate functional network measures into models of aging and AD, and to investigate environmental and lifestyle mediators linking education to brain network aging.

Older adults lacking a college degree show greater age-related decline in resting-state brain system segregation. Longitudinal decreases in segregation independently predict future increases in dementia severity, beyond APOE status, amyloid/tau pathology, and structural atrophy, and irrespective of education level. These findings identify functional network organization as a sensitive, prognostic indicator of brain health and preclinical cognitive decline not captured by structural measures or classical AD biomarkers. Future work should delineate the temporal ordering of network changes relative to clinical decline, integrate functional network measures into AD staging frameworks, and identify modifiable environmental and lifestyle factors mediating education-related disparities to inform prevention and intervention strategies.

- Methodological choices in graph analysis (parcellation, preprocessing, global signal regression, edge handling) can influence network measures; supplemental analyses suggest robustness, but further longitudinal validation is needed. - Sample composition after stringent QC underrepresents individuals with very low education and non-white participants; selection and attrition biases may underestimate effects, particularly among disadvantaged groups. - Health and lifestyle covariates were limited; detailed measures of nutrition, exercise, substance use, cognitive engagement, healthcare access/utilization, and early-life adversity were not available, limiting causal inference about mediators. - RSFC measures can be sensitive to head motion and arousal; extensive denoising and scrubbing were applied, yet residual confounding cannot be fully excluded. - Power may be insufficient to detect higher-order interactions among functional decline, APOE, pathology, and education in clinical prediction models.