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

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.

💡 Research Summary

The paper introduces a “neuro‑bridge” framework that tightly couples data‑driven spiking neural network (SNN) classification of resting‑state EEG with minimal biophysically grounded network simulations, aiming to provide both accurate Alzheimer’s disease (AD) detection and mechanistic insight into the underlying cortical circuitry. Clinical EEG recordings from 36 AD patients and 29 age‑matched healthy controls were pre‑processed (band‑pass filtering, artifact subspace reconstruction, ICA) and segmented into epochs. From each epoch, a set of interpretable features was extracted: conventional power‑band amplitudes, signal variance, phase‑locking value (PLV) based functional connectivity, and the aperiodic 1/f slope estimated with the FOOOF algorithm. The 1/f slope, known to reflect the excitation‑inhibition (E/I) balance at a macroscopic level, emerged as the most important predictor in SHAP‑based feature importance analysis.

These features were encoded into spike trains using a rate‑based scheme over 25 time steps and fed into a three‑layer SNN built with leaky‑integrate‑and‑fire neurons and surrogate gradients (implemented in snntorch). For comparison, an artificial neural network (ANN) with an identical dense architecture was trained on the same features. The SNN achieved an area under the ROC curve (AUC) of 0.839, comparable to the ANN, while offering the advantage of temporal spike‑based processing and a more direct mapping to excitatory and inhibitory neuronal populations.

To interpret the SNN’s reliance on the 1/f slope, the authors constructed minimal spiking network models in the NEST simulator. Each model comprised 400 neurons distributed across 19 cortical columns (one per EEG electrode). Within each column, three laminar groups (layers 2‑3, 4, and 5‑6) were represented, each containing an 80 % excitatory / 20 % inhibitory composition. By systematically varying the inhibitory‑to‑excitatory synaptic weight ratio, the simulations reproduced the empirical spectral slowing and reduced alpha power observed in AD. Two EEG proxies were examined: membrane potential and summed synaptic currents, both yielding consistent results.

Beyond local E/I manipulation, the authors incorporated empirical PLV‑derived functional connectivity as a prior in multi‑subnetwork simulations. This addition amplified spectral differentiation between simulated “healthy,” “MCI‑like,” and “AD‑like” states, indicating that large‑scale network topology exerts a stronger constraint on EEG signatures than E/I balance alone.

Overall, the study demonstrates three key contributions: (1) an energy‑efficient SNN classifier that leverages interpretable EEG features, (2) a quantitative link between the macroscopic 1/f slope and microscopic E/I ratio via biophysical modeling, and (3) evidence that functional connectivity priors enhance the explanatory power of circuit‑level simulations. By bridging machine learning and mechanistic modeling, the work moves EEG‑based AD diagnostics from opaque black‑box predictions toward explainable, neuroscience‑grounded biomarkers, and suggests a scalable pathway for neuromorphic implementation in clinical settings.