Scholarlink

<p>Alzheimer’s Disease (AD) is a progressive neurodegenerative, chronic disease that affects millions of people worldwide. EEG biomarkers are also an emerging and promising non-invasive tool for detecting early-stage AD. Existing models lack biological plausibility and robustness against noisy or incomplete signals. This study addresses these limitations by integrating graph-based spiking dynamics with domain-guided attention mechanisms. The proposed model, Neuro-SpikeNet, combines Dynamic Spiking Graph Neural Networks (DSGNN) and Knowledge-aware Fine-grained Attention Networks (KFAN) for precise AD classification. EEG data from two cohorts—one recorded with the Galileo BE Plus PRO system (35 subjects), the AFAVA dataset from Spain (23 subjects), and the BrainLat dataset from Latin America (780 subjects)—were used. Signals were preprocessed using the Iterative Robust Peak-aware Guided Filter (IRPGF) to suppress artifacts while preserving neural peaks. Feature extraction employed the Two-Sided Quaternion Windowed Quadratic-Phase Fourier Transform (TS-QW-QPFT) for detailed spectral and phase representation. Functional brain connectivity was modeled as dynamic graphs and encoded using Leaky Integrate-and-Fire (LIF) neurons. The KFAN employed biologically motivated attention over spatial–temporal regions involved in AD pathology. Hyperparameters were optimized using the Blood-Sucking Leech Optimizer (BSLO) to increase effectiveness. Neuro-SpikeNet achieved 99.99% accuracy in classification, with an AUC of 0.999 and a F1-score of 99.98%. The model performed well even when signals were noisy and data were incomplete. The architecture provides a biologically rational, scalable, and high-fidelity EEG-based approach to diagnose AD, with good applicability to early cognitive screening and clinical decision support systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neuro-SpikeNet: Optimized Spiking Graph and Knowledge-Aware Networks for High-Precision EEG-Based Alzheimer’s Diagnosis

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • T. Rajesh Kumar,
  • Krishna Prakash Arunachalam,
  • V. Deepak,
  • Elangovan Muniyandy

摘要

Alzheimer’s Disease (AD) is a progressive neurodegenerative, chronic disease that affects millions of people worldwide. EEG biomarkers are also an emerging and promising non-invasive tool for detecting early-stage AD. Existing models lack biological plausibility and robustness against noisy or incomplete signals. This study addresses these limitations by integrating graph-based spiking dynamics with domain-guided attention mechanisms. The proposed model, Neuro-SpikeNet, combines Dynamic Spiking Graph Neural Networks (DSGNN) and Knowledge-aware Fine-grained Attention Networks (KFAN) for precise AD classification. EEG data from two cohorts—one recorded with the Galileo BE Plus PRO system (35 subjects), the AFAVA dataset from Spain (23 subjects), and the BrainLat dataset from Latin America (780 subjects)—were used. Signals were preprocessed using the Iterative Robust Peak-aware Guided Filter (IRPGF) to suppress artifacts while preserving neural peaks. Feature extraction employed the Two-Sided Quaternion Windowed Quadratic-Phase Fourier Transform (TS-QW-QPFT) for detailed spectral and phase representation. Functional brain connectivity was modeled as dynamic graphs and encoded using Leaky Integrate-and-Fire (LIF) neurons. The KFAN employed biologically motivated attention over spatial–temporal regions involved in AD pathology. Hyperparameters were optimized using the Blood-Sucking Leech Optimizer (BSLO) to increase effectiveness. Neuro-SpikeNet achieved 99.99% accuracy in classification, with an AUC of 0.999 and a F1-score of 99.98%. The model performed well even when signals were noisy and data were incomplete. The architecture provides a biologically rational, scalable, and high-fidelity EEG-based approach to diagnose AD, with good applicability to early cognitive screening and clinical decision support systems.