Explanatory limitations of CSF, plasma, and MRI biomarkers for inflammation and blood–brain barrier disruption in dementia
TL;DR
The Qalb–CSF model showed the strongest inflammation-related signal, but these findings are exploratory and require validation in independent, longitudinal cohorts before an inflammation axis can be incorporated into an ATN(V)-based framework.
Abstract
Inflammation and blood–brain barrier (BBB) disruption are increasingly implicated in cognitive decline, but it remains unclear which fluid and imaging measures best capture these processes. Participants were drawn from two overlapping research programs, MarkVCID and UNM ADRC. We studied 149 participants with Alzheimer’s disease, leukoaraiosis, mixed dementia, subcortical ischemic vascular dementia, or memory impairment. BBB permeability was assessed using the albumin index (Qalb), a global blood-to-CSF leakage ratio, and dynamic contrast-enhanced MRI permeability (Ktrans), a regional gadolinium transfer measure. Both were treated as permeability measures that inflammation may influence rather than as direct measures of inflammation. CSF and plasma biomarkers included matrix metalloproteinases, angiogenic factors, cytokines, GFAP, NfL, pTau181, and Aβ42/40; MRI measures included PSMD, mean free water, hippocampal volume, and cortical thickness. Univariate associations were screened using Spearman correlation and AIC, and multivariable models used AIC-based stepwise selection. Diagnosis-adjusted models, bootstrap selection frequencies, penalized regression, and cross-validated R2 were used as sensitivity analyses. Qalb and Ktrans were not significantly correlated (Spearman ρ = 0.14, p = 0.28), suggesting that they index different dimensions of BBB dysfunction without establishing biological independence. The multivariable CSF model for Qalb retained MMP-2, Flt-1, IL-8, GFAP, and NfL and explained approximately half the variance (R2 = 0.53), whereas the plasma model explained less variance (R2 = 0.32). For Ktrans, multivariable CSF, plasma, and MRI models explained R2 = 0.36, 0.39, and 0.14, respectively, while PSMD was the strongest FDR-robust univariate MRI correlate. CSF–plasma correlations ranged from near zero for MMP-9 (r = −0.03) to stronger associations for NfL (r = 0.74) and IL-13 (r = 0.70), indicating that plasma cannot be assumed to substitute for CSF marker-by-marker. Qalb and Ktrans should be considered complementary rather than interchangeable measures. The Qalb–CSF model showed the strongest inflammation-related signal, but these findings are exploratory and require validation in independent, longitudinal cohorts before an inflammation axis can be incorporated into an ATN(V)-based framework.