Abstract
Alzheimer's disease (AD) is a heterogeneous disorder characterized by brain accumulation of amyloid-beta (Aß, simplified as A for the AD biological model) and tau (T) proteins, with Aß emerging first. However, a significant proportion of individuals exhibit discordant biomarkers' profiles, such as elevated phosphorylated tau181 (p-tau181) with normal Aß42 from cerebrospinal fluid (CSF), posing diagnostic and mechanistic challenges. This study investigated whether functional and structural brain connectivity can distinguish individuals with discordant CSF profiles (A-T+) from those with concordant patterns (A+T+), hypothesizing that distinct connectivity patterns may reflect early divergent pathophysiological processes. Data from cognitively unimpaired or mildly impaired individuals in the ADNI3 repository were analyzed, selecting those with resting-state functional MRI (rsfMRI) and/or diffusion MRI (dMRI) within 18 months of CSF testing for Aß and p-tau181. Participants were grouped into A-T+ or A+T+ groups. Structural and functional connectivity gradients were generated for each participant and summarized using a Euclidean distance measure from reference gradient templates derived from cognitively unimpaired individuals without pathology (A-T-). We applied linear mixed models and analysis of variance to assess connectivity-based gradient differences between A-T+ and A+T+ groups, adjusting for relevant variables. Classification analyzes using logistic regression and support vector machine, along with feature importance via the Boruta algorithm, evaluated the discriminative power of gradient connectivity profiles. Multimodal integration was performed using partial least square canonical analysis (PLSC), and relationships between gradients and cognition were assessed via UMAP-based dimensionality reduction and bootstrapped linear regressions. Results were compared with a classical network analysis examining within- and between-network connectivity differences. Among 424 participants, n = 67 were classified as A-T+, n = 106 as A+T+, and n = 56 as cognitively healthy A-T-. The remaining 195 participants (n = 86 A+T+ and n = 109 cognitively impaired A-T-) were not included. A-T+ individuals (age = 75 ± 8.2) exhibited less cognitive impairment but greater functional connectivity gradients' distance to the reference templates (false discovery rate-corrected p < 0.05) in the temporo-occipital axis compared to A+T+ (age = 76.1 ± 7.7). Structural connectivity differences were not significant. FC-based models classified A-T+ and A+T+ with good accuracy (AUC = 0.77), loading on the same temporo-occipital regions, unlike SC (AUC = 0.52). The posterior brain involvement in A-T+ was confirmed by PLSC analyzes. A+T+ individuals showed a significant relation between cognitive scores and functional connectivity, primarily mapping the default mode network (DMN). A shift was observed in relation to executive functions and functional connectivity in A-T+. Discordant CSF profiles (A-T+) exhibit distinct functional connectivity patterns, particularly in posterior brain regions, compared to concordant CSF patterns (A+T+), which are characterized by a significant cognitive-DMN connectivity association. These results suggest that CSF p-tau181 accumulation in the absence of Aß42 may be associated with specific functional trajectories, suggesting specific pathophysiological patterns.