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

The paper proposes a comprehensive framework for identifying robust signatures of Alzheimer's Disease (AD) by integrating electroencephalography (EEG) data with biophysical simulations utilizing spiking neural networks. This approach is critical because AD pathology manifests across multiple levels—from synaptic dysfunction to large-scale network breakdown—requiring a multiscale methodology that moves beyond traditional spectral analysis. By bridging empirical physiological measurements with computational models, the research aims to provide a unified approach to interpreting model predictions and develop more sensitive, early diagnostic biomarkers.

Biophysical Modeling of Neuronal Dynamics

The foundation of this investigation rests on modeling neuronal activity using spiking neural networks (SNNs) and integrate-and-fire models. These computational tools allow researchers to simulate how complex biological processes occur at the cellular level, such as signal propagation and logic gating in networks. Key aspects modeled include:

• Excitation/Inhibition Balance: Methods are employed for inferring synaptic excitation/inhibition balance from field potentials, recognizing that imbalances are central to AD pathophysiology.

• Network Dynamics: The dynamics of sparsely connected networks, particularly those involving excitatory and inhibitory spiking neurons, are simulated to understand how structural changes impact overall network function.

Advanced EEG Spectral and Aperiodic Analysis

Traditional spectral analysis is enhanced by methods that decompose power spectra into distinct components. Researchers utilize techniques for parameterizing neural power spectra into periodic and aperiodic components. The aperiodic component, specifically the 1/f slope, is highlighted as a potential marker of underlying pathology. Furthermore, longitudinal studies have established that changes in EEG patterns are observable over time; for instance, research tracks longitudinal EEG study[ies] of mild senile dementia of Alzheimer type, noting specific spectral shifts associated with cognitive decline.

Quantifying Network Connectivity and Synchronization

Identifying pathological signatures requires moving beyond single-channel analysis to examine connectivity patterns. Several quantitative metrics are employed to map network alterations:

• Network Statistics: Techniques like the network-based statistic are used for identifying differences in brain networks between healthy and diseased states.

• Phase Consistency: The pairwise phase consistency serves as a bias-free measure designed to quantify rhythmic neuronal synchronization, which is expected to degrade in AD.

• Functional Disruption: Studies examine how disrupted brain functional connectivity... [serves as] early signature in cognitively healthy individuals with pathological csf amyloid/tau.

Multimodal Integration for Biomarker Discovery

The most advanced aspect of the methodology involves fusing data from disparate sources—EEG, fNIRS, and biophysical simulations—to create holistic diagnostic profiles. This integration allows for the assessment of complex interactions:

• Cross-Modal Analysis: Investigations are conducted on investigating the interaction between EEG and fnirs: A multimodal network analysis of brain connectivity.

• Combined Biomarkers: The framework supports combining physiological markers, such as aperiodic 1/f slopes, with simulation results to provide a refined proxy marker for metabolic imbalance.

• Whole-Brain Modeling: Ultimately, the goal is to create comprehensive models that use diverse receptor maps and functional connectivity data to explain non-linear effects observed across the entire brain system.

(Self-Correction Protocol Initiated: Due to the high stakes, I must first address the input limitation. Since only a bibliography is provided and not the full arXiv paper text, I must assume that this entire set of references defines the core scientific domain of the missing manuscript. Therefore, my improvements will focus on bridging advanced computational neuroscience principles (as derived from these citations) with state-of-the-art, high-reliability AI architectures.)

The current research domain excels at describing neural dysfunction (e.g., measuring reduced alpha power [29], identifying altered functional connectivity [16], or quantifying E/I imbalance [25]). The primary limitation for clinical deployment, and thus the major AI opportunity, is moving from correlation-based diagnosis to causally informed, longitudinal prediction.

Here are three critical improvements that must be integrated into the AI pipeline:

Current methods treat connectivity metrics (functional, phase consistency [20]) and spectral features (gamma band power [32], 1/f slope [21]) as separate inputs. This is computationally inefficient and biologically incomplete.

• The Improvement: Implement a unified Graph Neural Network (GNN) architecture that dynamically constructs the brain connectome graph at multiple temporal and frequency resolutions simultaneously.

• Mechanism: The GNN must accept heterogeneous node features (e.g., spectral power from [17], local E/I estimates from [25]) and edge weights derived from different connectivity metrics (e.g., Phase Locking Value, Phase Consistency [20]). This requires a Graph Attention Network (GAT) layer to learn the relative importance of different connection types (e.g., is phase consistency more informative than spectral coherence for MCI?).

• What the Improved AI System Can Do: It can build a single, robust representation of the patient's network state that simultaneously encodes where connectivity is lost (structural nodes), how it is changing over time (temporal dynamics), and what frequency band is affected (spectral features). This drastically improves feature robustness compared to concatenating disparate metrics.

Most current models are classifiers (State to Diagnosis). For clinical cost-saving, we need predictors (Current State + Trajectory to Future Event). The underlying mechanism of AD is a process of decline, not a static endpoint.

• The Improvement: Integrate Causal Discovery Algorithms (e.g., based on Transfer Entropy or advanced state-space models like Variational Autoencoders adapted for causality) into the GNN output layer.

• Mechanism: Instead of training the model to predict P(AD EEG), train it to estimate the directional flow of pathology. The AI must learn directional dependencies: Does reduced Theta power in Region A precede reduced gamma band activity in Region B? This allows the system to identify leading indicators (precursors) rather than merely co-occurring biomarkers.

• What the Improved AI System Can Do: It can output a probabilistic risk trajectory, estimating not just if a patient has MCI, but the likelihood and estimated timeframe of progression to AD (P(AD within 3 years)). This is essential for optimizing intervention timing.

Given the high cost of errors, a black box prediction is unacceptable. The AI must provide biologically plausible justification for its prediction based on established neurophysiology principles derived from the literature (e.g., E/I balance, network hub theory).

• The Improvement: Implement Physics-Informed Neural Networks (PINNs) or attention mechanisms constrained by known biological laws.

• Mechanism: The loss function must be augmented with a regularization term that penalizes predictions violating fundamental neurophysiological principles derived from the literature (e.g., if the model predicts severe network failure, it must also predict a corresponding collapse in local E/I balance metrics). Furthermore, Attention Weights must be mapped back to specific biological nodes or pathways (e.g., The prediction is weighted most heavily by the disruption of the Default Mode Network hub connectivity).

• What the Improved AI System Can Do: It provides Actionable Interpretability. Instead of stating, Diagnosis: High Risk, it states: "Diagnosis: High Risk (Confidence=92%). Primary driving factor: Reduced coupling strength in the Posterior Cingulate Cortex pathway (Citation [27] suggests this). Recommended intervention focus: Target restoring local excitation/inhibition balance in this specific network module." This moves the AI from a diagnostic tool to a personalized therapeutic guidance system.

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.