Learning Alzheimer's Disease Signatures by bridging EEG with Spiking Neural Networks and Biophysical Simulations

TL;DR

Alzheimer's disease EEG biomarkers are challenging to interpret and deploy; this paper bridges data-driven neuromorphic learning with mechanistic biophysical simulations to connect macroscopic EEG signatures to cortical circuit dynamics. The authors show an SNN classifier achieving an AUROC of $0.839$ on resting-state EEG and identify the aperiodic $1/f$ slope as a key discriminant, while simulations across E/I balance reproduce spectral changes observed in AD. They further demonstrate that incorporating band-specific functional connectivity priors sharpens spectral differentiation and reduces overestimation of E/I effects, highlighting the importance of large-scale network topology. Collectively, the neuro-bridge advances mechanistic interpretability and paves the way for scalable, explainable, energy-efficient EEG-based AD screening.

Abstract

As the prevalence of Alzheimer's disease (AD) rises, improving mechanistic insight from non-invasive biomarkers is increasingly critical. Recent work suggests that circuit-level brain alterations manifest as changes in electroencephalography (EEG) spectral features detectable by machine learning. However, conventional deep learning approaches for EEG-based AD detection are computationally intensive and mechanistically opaque. Spiking neural networks (SNNs) offer a biologically plausible and energy-efficient alternative, yet their application to AD diagnosis remains largely unexplored. We propose a neuro-bridge framework that links data-driven learning with minimal, biophysically grounded simulations, enabling bidirectional interpretation between machine learning signatures and circuit-level mechanisms in AD. Using resting-state clinical EEG, we train an SNN classifier that achieves competitive performance (AUC = 0.839) and identifies the aperiodic 1/f slope as a key discriminative marker. The 1/f slope reflects excitation-inhibition balance. To interpret this mechanistically, we construct spiking network simulations in which inhibitory-to-excitatory synaptic ratios are systematically varied to emulate healthy, mild cognitive impairment, and AD-like states. Using both membrane potential-based and synaptic current-based EEG proxies, we reproduce empirical spectral slowing and altered alpha organization. Incorporating empirical functional connectivity priors into multi-subnetwork simulations further enhances spectral differentiation, demonstrating that large-scale network topology constrains EEG signatures more strongly than excitation-inhibition balance alone. Overall, this neuro-bridge approach connects SNN-based classification with interpretable circuit simulations, advancing mechanistic understanding of EEG biomarkers while enabling scalable, explainable AD detection.

Figures (9)

• Figure 1: Simplistic overview of the connections between SNN classifier and minimal biophisical simulation.
• Figure 2: SNN-based EEG classification pipeline. Raw EEG epochs (19 channels × 500 samples), with each epoch/window having 500 time points. Those then undergo feature extraction to compute spectral power bands, phase-locking values (PLV), and aperiodic slope (1/f). The resulting 200-dimensional feature vectors are batched (B=128), encoded into spike trains via rate-based encoding over T=25 time steps, and classified by a 3-layer spiking neural network with leaky integrate-and-fire neurons. The network outputs class predictions C (AD vs. HC) for each epoch in the batch. For the ANN classifier the pipeline is the same except the spike train encoding.
• Figure 3: ROC for both ANN (green line) and SNN (gray line) classifier
• Figure 4: Top 10 global feature relevance analysis of the SNN and ANN using SHAP. Each subfigure represents the mean SHAP values across EEG windows for SNN (left) and ANN (right). Each dot represents one EEG sample / epoch / subject evaluated for that feature. 200 Multiple EEG-derived features were used, including spectral bands, connectivity metrics, standard deviation (signal amplitude variability) and the inverse power-law slope of the EEG spectrum.
• Figure 5: Significant functional connectivity edges identified by the NBS analysis depicted in axial views. Subfigures (a–e) show the significant connections for the left-tailed NBS contrast in the delta, theta, alpha, beta, and gamma frequency bands, respectively. Subfigures (f–j) report the corresponding results for the right-tailed NBS contrast across the same frequency bands. Each subfigure displays only the edges belonging to the significant NBS component detected for that frequency band; subfigures with no visible edges indicate that no significant component was found. Subfigure (k) shows the EEG electrode layout, provided as a spatial reference for interpreting the connectivity patterns.