Neural alterations associated with Alzheimer’s disease (AD) may emerge years before clinical diagnosis, and genetic risk factors such as APOE-ε4 and PICALM rs3851179 may identify individuals vulnerable to these shifts. We tested whether electroencephalography (EEG) signatures learned from clinically diagnosed AD patients transfer to cognitively healthy middle-aged adults carrying these variants. Differentiable logic gate networks (Diff-Logic) and neural baselines were trained on log-relative power spectral density features from a clinical cohort of AD patients and cognitively normal (CN) controls. These models were then applied without retraining to an independent genetic-risk cohort of neutral-risk non-carriers (N), APOE-ε4-only carriers (A+P–), and dual APOE-ε4/PICALM G-allele carriers (A+P+), yielding a continuous AD-like EEG signature score per participant. Diff-Logic matched neural baselines in clinical classification (balanced accuracy 0.809±0.028; AUROC 0.845±0.023) while enabling edge inference with a 7.4× model-size reduction and 0.22 ms model-only latency on a 7 W single-core CPU. In the genetic-risk cohort, scores were highest in A+P+ participants (0.393±0.167) relative to N (0.296±0.154) and A+P– (0.285±0.143; subject-level mean ± SD). This elevation was nominally significant before correction (A+P+ vs. A+P–, punc=0.036) but did not survive Holm correction (pHolm=0.109), and the cohort carried no AD pathology or conversion ground truth; the association was therefore exploratory. Integrated gradients attribution of the trained model showed genetic-risk relevance in delta, theta, and alpha, partially overlapping clinical AD topography. These findings support the feasibility of zero-shot EEG transfer in genetically at-risk individuals and highlight the need for replication in larger cohorts. Code is publicly available at https://github.com/Shyamal-Dharia/ad-difflogic-transfer.
