Alzheimer's EEG Binary Classifier
Comparing architectural assumptions for EEG classification
A comparative study of three deep-learning architectures — EEGNet (compact convolutional), EEGTransformer (self-attention), and EEGGAT (graph attention) — for binary classification of resting-state EEG into Normal vs Impaired (MCI + moderate/severe AD).
- EEG
- EEGNet
- Transformer
- GAT
- Alzheimer's
- Resting-State
- PyTorch
Question
Why It Matters
System
Data
Method
Preprocessing
Bandpass filter 0.5–45 Hz, notch filter 50 Hz, resample to 250 Hz, 2-second non-overlapping epochs (500 timepoints), z-score normalization per channel. Identical for all three models.
Feature Extraction
Learned features within each architecture — no shared manual feature extraction step. EEGNet learns local temporal convolutions; Transformer learns global self-attention; GAT learns graph-attention over electrode relationships.
Model
EEGNet (compact convolutional, Lawhern et al. 2018), EEGTransformer (self-attention over temporal sequences), EEGGAT (graph attention over 10-20 electrode adjacency).
Training
Subject-wise stratified split (70% train, 15% val, 15% test). Focal loss and data augmentation (noise, channel dropout, time shift) to prevent overfitting. Each model trained independently under matched conditions for fair comparison.
Evaluation
Test set accuracy, balanced accuracy (val), precision/recall/F1 for the Impaired class. Subject-wise splitting is mandatory — epoch-wise splits give overly optimistic results.
Experiment
Results
| Model | Test Acc | Balanced Acc (Val) | Precision (Impaired) | Recall (Impaired) | F1 (Impaired) |
|---|---|---|---|---|---|
| EEGNet | 66.0% | 72.8% | 0.88 | 0.62 | 0.73 |
| EEGTransformer | 74.1% | 75.5% | 0.79 | 0.88 | 0.83 |
| EEGGAT | 65.8% | 63.8% | 0.84 | 0.66 | 0.74 |
Test set results. Subject-wise stratified split (70/15/15). Focal loss + augmentation used for all models.
The Transformer model achieved the highest test accuracy (74.1%) and excellent recall for the Impaired class (0.88), making it the most clinically useful for screening — high recall means fewer missed cases. EEGNet had higher precision for Impaired (0.88) but lower recall (0.62), meaning fewer false positives but more false negatives. The GAT underperformed (65.8%), likely because the static 10-20 adjacency graph provides limited connectivity information with only 19 channels. Focal loss and data augmentation were critical to prevent overfitting across all models.
What Changed
The original dataset contains three groups (Alzheimer's, Frontotemporal dementia, healthy controls); MCI and moderate/severe AD were merged into a single "Impaired" class to create a clinically relevant binary screening task. Subject-wise splitting was found to be mandatory — epoch-wise splits give overly optimistic results because epochs from the same subject leak into both train and test sets.
Limitations
- Research/experimental project — not a medical diagnostic tool. No clinical validation of any kind.
- Dataset-specific results on one dataset (DS007526) — do not generalize automatically to other populations or recording setups.
- The GAT's static 10-20 adjacency graph may be too simplistic — future work could use data-driven connectivity.
- Binary task only — the original three-group distinction (AD vs FTD vs healthy) was collapsed.
- Comparative evaluation is offline only — not a real-time screening system.
Next Experiment
Analyze per-class failure modes across architectures and test whether data-driven graph connectivity (instead of static 10-20 adjacency) improves GAT performance. Then apply the same methodology to prodromal Parkinson's disease detection using the TDBrain dataset.