Youssef Tag Eldin

Youssef Tag Eldin

Machine Learning Engineer with experience in deep learning, classical ML, and applied AI

EEG-Based Alzheimer’s Disease Classification

Project Overview

Project Pipeline
End-to-end pipeline from EEG signal acquisition to disease classification

This project presents a novel graph-based transformer architecture for the classification of Alzheimer’s Disease and Frontotemporal Dementia using EEG signals. Developed as my Bachelor’s thesis, this work achieved 99.1% accuracy, significantly surpassing existing state-of-the-art methods by 7-18 percentage points.

Key Achievements

MetricResult
Classification Accuracy99.1%
Improvement over SOTA7-18 percentage points
ClassesHealthy, Alzheimer’s, Frontotemporal Dementia
StatusUnder review at Journal of Neural Engineering

Model Architecture

Model Architecture
Graph-based transformer architecture integrating spatial–temporal EEG features

The framework integrates spatial and temporal EEG features using a custom graph-transformer design:

EEG Feature Visualization

Our approach leverages power spectral density (PSD) features across EEG frequency bands:

EEG Brain Heatmap PSD Comparison
Left: Power spectral density (PSD) heatmap across brain regions. Right: Spectral analysis showing differences between healthy controls, Alzheimer’s, and FTD patients.

The analysis highlights distinct biomarkers distinguishing the three groups. The spatial heatmaps reveal localized changes in brain activity intensity, while the spectral plots demonstrate clear separation in power distribution frequencies—specifically showing how Alzheimer’s pathology alters lower-frequency band power compared to healthy controls.

Technical Implementation

  • Graph Neural Networks: Custom graph transformer leveraging electrode spatial relationships
  • Feature Engineering: Relative band power, spectral coherence across EEG frequency bands
  • Signal Processing: Comprehensive EEG preprocessing and artifact removal
  • Benchmarking: Systematic comparison of classical ML vs. deep learning approaches

Technologies Used

  • Python, PyTorch, PyTorch Geometric
  • scikit-learn, MNE-Python
  • Signal processing: FFT, spectral analysis

Research Impact

This work demonstrates the potential of graph-based deep learning for biomedical signal processing. The significant improvement over existing methods suggests new directions for non-invasive Alzheimer’s diagnostics using routine EEG recordings.

Publication Status

Under Review at IOP Journal of Neural Engineering
First author — responsible for design, analysis, and drafting

Code Repository

github.com/yosftag1/eeg-alzheimer-classification

Project Timeline

September 2023 – June 2024 (Bachelor’s Thesis Project)
Supervised by: Dr. A. Eltrass, Dr. H. Farag