Abstract
<title>Abstract</title> <p>Early, scalable detection of Alzheimer's disease (AD) and mild cognitive impairment (MCI) remains a global health priority, but non-invasive biomarkers that capture network-level dysfunction during standardized cognitive challenge remain limited. We coupled immersive virtual reality (VR)-synchronised 64-channel electroencephalography (EEG) with phase transfer entropy (PTE), a directional connectivity measure, to identify interpretable electrophysiological signatures across the AD continuum. In a quality-controlled cohort of 37 participants (AD n = 13, MCI n = 13, normal cognition n = 11; 72 recruited, 60 completed the protocol), 16,892 features across five computational domains were evaluated using a six-classifier pipeline with leave-one-out cross-validation, permutation testing, bootstrap resampling, and SHAP-based interpretation. The best-performing model achieved 86.5% three-class accuracy with 100% MCI recall (macro-F1 = 0.859) and identified 11 candidate biomarkers supported by cross-validated stability and effect-size evidence. Key signatures included attenuated posterior alpha net flow in AD, VR-elicited parieto-occipital gamma connectivity in MCI, and preserved central gamma hub structure in normal cognition. These findings suggest that VR-EEG directional connectivity provides a reproducible framework for identifying candidate electrophysiological biomarkers of the AD continuum, warranting validation in larger biomarker-confirmed cohorts.</p>