Resting-state fMRI can reveal network alterations in Alzheimer’s disease, but connectivity-based classification is challenged by high dimensionality, acquisition heterogeneity, and unstable interpretation.

Objective: To evaluate a leakage-controlled latent-modelling framework for CN–AD classification and the identification of stable, anatomically interpretable connectivity signatures.

rs-fMRI acquisition (ADNI): 3T EPI · TR 3 s · TE 30 ms · 7–10 min (140/200 volumes) · Philips / Siemens / GE.

Preprocessing: DPARSF with motion correction, MNI normalisation, 4-mm smoothing, 0.01–0.10 Hz band-pass, nuisance regression.

Parcellation: AAL3 → 131 ROIs, Yeo-17 ordered, 140 z-scored timepoints.

External validation: OASIS, N=180 (90 CN / 90 AD), TR 2.5 s, 164 volumes.

A convolutional β-variational autoencoder learns a compact, diagnosis-agnostic latent embedding of multichannel functional connectivity (3 selected channels). Latent means with age and sex feed a strictly nested 5×5 cross-validated classifier. SHAP and Integrated Gradients map fold-stable signal back to connectivity edges.

Classification performance

L2 Logistic Regression applied to the 384 dimensions β-VAE latent features, together with age and sex, provided the strongest CN/AD discrimination, well above the 24.4% AD-prevalence baseline (N = 397; 97 AD / 300 CN).

External OASIS evaluation: ROC–AUC 0.648; PR–AUC 0.668. PR baseline = 0.50.

Interpretability

SHAP + Integrated Gradients + cross-fold consensus: a compact 8-edge signature (6 pro-AD, 2 pro-CN) spanning default-mode, control, dorsal-attention, limbic and somatomotor systems.

Top edge: Parietal_Sup_R – Precuneus_R (pro-CN).

Conclusions