Pre-symptomatic Alzheimer’s risk detection is difficult precisely because genetically at-risk individuals look neurologically normal for years before any clinical presentation. This paper presents a multimodal neurophysiological framework distinguishing two biologically distinct risk pathways, APOE ε4 (N = 48) and PICALM rs3851179 (N = 30), using resting-state EEG and fMRI in a pre-symptomatic cohort (µage ≈55 years) from the PEARL-Neuro dataset (OpenNeuro ds004796). EEGNet trained under leave-one-subject-out cross-validation reveals a bimodal fold structure in which 57.8% of APOE e3/e4 carriers and 80.0% of PICALM carriers are correctly classified at the subject level, with no intermediate folds between near-perfect (F1 = 1.0, 50/77) and near-chance (F1 ≈0.47–0.50, 27/77) performance. This structure reflects genuine individual variability in pre-symptomatic EEG divergence onset, with age confirmed as non-explanatory (Mann-Whitney U, p>.05). SHAP analysis and EEGNet’s learned spatial filters converge independently on bilateral temporal and frontal channels as the primary discriminating anatomy. Default Mode Network (DMN) connectivity is the single most discriminative individual feature (mean |SHAP|= 0.038) despite showing no linear correlation with any EEG measure (r .16, N = 68), indicating the two modalities access distinct biological aspects of the same risk process. Late fusion demonstrates a ceiling effect consistent with EEGNet’s near-perfect unimodal performance, but within the 24 maximally uncertain subjects, DMN provides genuine probability calibration improvement of approximately 0.20–0.25 in P(PICALM risk). A companion ResNet18 analysis on OASIS-2 (N = 150 baseline sessions) contributes neuroanatomical grounding through Grad-CAM, identifying posterior-inferior atrophy as the primary CDR-based MCI discriminator. Together, these results suggest the preclinical Alzheimer’s neural phenotype is modality-specific, pathway-dependent, and individually variable in its timing of emergence.
Shlok Khare (Tue,) studied this question.