Stable Network-Level Functional Connectivity Alterations in Alzheimer's Disease Identified via Interpretable Latent Modelling

Resting-state fMRI provides a non-invasive window into large-scale network-level alterations in Alzheimer's disease (AD), but the high-dimensional functional connectivity (FC) and multi-site heterogeneity pose challenges to both classification and interpretability. We propose an explainable deep-learning framework that combines diagnosis-agnostic latent representation learning with a rigorously nested and interpretable classification pipeline to identify reproducible connectivity biomarkers of AD. Using a multi-site ADNI cohort (rs-fMRI, N=431; 95 AD, 89 CN, 247 MCI), a diagnosis-agnostic b-VAE was trained diagnosis-agnostically on CN+MCI+AD data to learn a smooth latent representation of multichannel connectivity. We evaluated seven candidate connectivity measures spanning static, graph-filtered, dynamic, and effective connectivity. A systematic ablation study identified three complementary static channels as the most informative: Full Pearson correlation, OMST-reweighted Pearson, and k-NN-based Mutual Information. Fold-specific latent means were then combined with age and sex and classified under a strictly nested 5 x 5 cross-validation scheme. Across outer test folds, Logistic Regression achieved a mean ROC--AUC of 0.843 and PR--AUC of 0.864. Pooled out-of-fold predictions for CN vs. AD subjects yielded a ROC--AUC 0.829 and a low Brier score (0.197), indicating well-calibrated probabilities. An attribution pipeline combining SHAP in latent space, followed by Integrated Gradients backprojected to edges, recovered a compact and reproducible AD signature. At the systems level, the differential saliency highlights Default Mode, Limbic, and Visual networks. Enforcing strict croos-fold stability criteria (replication frequency and sign consistency) yielded a consensus subgraph of 11 highly reliable connections. Importantly, the identified multivariate signature was largely disjoint from the strongest mass-univariate effects, indicating that the model captures complementary information beyond marginal group differences. Confounder analyses showed only moderate persistence of scanner-related signal in the latent space, with no evidence that acquisition or demographic factors systematically drove AD classification performance. Together, these results demonstrate a principled route toward interpretable, reproducible, and well-calibrated connectivity-based biomarkers for Alzheimer's disease. All code, trained models, and mappings are publicly available to support transparency and external validation.