EEG-GCNN: Dataset, Task, Models, and Neurological Disease Detection Pipeline

[jburhan]

Contributors

• Jimmy Burhan — jburhan2 (Dataset + Task)
• Robert Coffey — rc37 (Models + Training)

Type of Contribution

Option 4: Full Pipeline — Dataset + Task + Model (60 pts)

Paper

Wagh & Varatharajah (2020) — EEG-GCNN: Augmenting Electroencephalogram-based Neurological Disease Diagnosis using a Domain-guided Graph Convolutional Neural Network, ML4H @ NeurIPS 2020

Summary

End-to-end EEG-GCNN pipeline for neurological disease detection from EEG recordings, replicating and extending Wagh & Varatharajah (2020).

Dataset & Task (Jimmy Burhan — jburhan2)

• EEGGCNNRawDataset — raw EEG loader for TUAB (1,385 clinically-normal patients, label 0) + LEMON (208 healthy volunteers, label 1): resamples to 250 Hz, 1 Hz high-pass + 60/50 Hz notch filter, 8 bipolar channels, 10 s non-overlapping windows, 6-band Welch PSD features (delta/theta/alpha/beta/low_gamma/high_gamma), writes 5 FigShare-format arrays via precompute_features()
• EEGGCNNDataset — FigShare pre-computed dataset (1,593 subjects, 225,334 windows); builds graph adjacency via tunable blend of geodesic distance and spectral coherence: edge_weight = α·geodesic + (1−α)·coherence
• EEGGCNNDiseaseDetection — per-sample streaming task with configurable adjacency type (spatial / functional / combined) and band subset; drives both ablation studies
• EEGGCNNClassification — window-level classification task for FigShare data

Models & Training (Robert Coffey — rc37)

• EEGGraphConvNet — two-layer GCN classifier (reproduces paper's Shallow EEG-GCNN, float32)
• EEGGATConvNet — two-layer graph attention network classifier (float64)
• Training pipelines, holdout evaluation scripts, band frequency ablation script
• 23 unit tests (synthetic fixtures, no real EEG required)

Ablation Studies (Original Extensions)

Ablation 1 — Adjacency Type (α sweep, GCN)

Varied α ∈ {0.0, 0.25, 0.5, 0.75, 1.0} on the full FigShare dataset (477 test patients, 70/30 split, 10-fold CV). Paper only uses α=0.5 — this is our novel contribution:

α	auroc_patient	bal_acc	f1
0.0 (coherence only)	0.8841 ± 0.0132	0.8087 ± 0.0130	0.8222 ± 0.0176
0.25	0.8943 ± 0.0072	0.8069 ± 0.0122	0.8385 ± 0.0162
0.50 (paper default)	0.8984 ± 0.0102	0.8276 ± 0.0122	0.8399 ± 0.0281
0.75	0.8991 ± 0.0066	0.8209 ± 0.0051	0.8491 ± 0.0323
1.0 (geodesic only)	0.7891 ± 0.0471	0.7406 ± 0.0349	0.8211 ± 0.0599

α=1.0 (spatial-only) collapses performance — functional connectivity is essential. α=0.75 best for auroc_patient and F1; α=0.50 best for balanced accuracy — both within each other's error margin.

Ablation 2 — Spectral Frequency Analysis (GCN, inference-only)

Leave-one-out (LOO) and keep-one-in (KOI) across all 6 bands using trained GCN checkpoints — no retraining required. Low gamma is the single most important band (removing it drops AUROC from 0.898 → 0.594; keeping only it recovers 0.723). All bands make a meaningful contribution — none is redundant. Numerical results in project slides.

Best Results (GAT, tuned learning rate)

auroc_patient	bal_acc	f1	precision	recall
0.9415 ± 0.0251	0.8854 ± 0.0322	0.9201 ± 0.0231	0.9855 ± 0.0089	0.8638 ± 0.0409

GAT with tuned LR outperforms GCN on every patient-level metric. End-to-end run on locally pre-computed features matched FigShare-trained results.

File Guide

File	Description
pyhealth/datasets/eeg_gcnn.py	EEGGCNNDataset — FigShare pre-computed loader
pyhealth/datasets/eeg_gcnn_raw.py	EEGGCNNRawDataset — raw TUAB + LEMON loader with precompute_features()
pyhealth/datasets/configs/eeg_gcnn.yaml	FigShare dataset config
pyhealth/datasets/configs/eeg_gcnn_raw.yaml	Raw dataset config
pyhealth/tasks/eeg_gcnn_classification.py	EEGGCNNClassification — window-level task for FigShare path
pyhealth/tasks/eeg_gcnn_disease_detection.py	EEGGCNNDiseaseDetection — streaming task for raw path
pyhealth/models/eeg_gcnn.py	EEGGraphConvNet (GCN)
pyhealth/models/eeg_gatcnn.py	EEGGATConvNet (GAT)
examples/eeg_gcnn/pre_compute.py	Batch precompute: raw TUAB + LEMON → 5 FigShare-format arrays
examples/eeg_gcnn/training_pipeline_shallow_gcnn.py	GCN training (10-fold CV)
examples/eeg_gcnn/heldout_test_run_gcnn.py	GCN holdout evaluation (Youden's J threshold)
examples/eeg_gcnn/run_band_ablation.py	Spectral frequency ablation (LOO + KOI, 13 conditions)
examples/eeg_gatcnn/pre_compute_gatcnn.py	Batch precompute for GAT (float64)
examples/eeg_gatcnn/training_pipeline_shallow_gatcnn.py	GAT training (tuned LR)
examples/eeg_gatcnn/heldout_test_run_gatcnn.py	GAT holdout evaluation
pyhealth/tests/test_eeg_gcnn.py	23 unit tests: dataset, task, models (all synthetic data)
docs/api/eeg_gcnn_pipeline.rst	Full pipeline documentation
docs/api/datasets/pyhealth.datasets.EEGGCNNDataset.rst	API doc — FigShare dataset
docs/api/datasets/pyhealth.datasets.EEGGCNNRawDataset.rst	API doc — raw dataset
docs/api/models/pyhealth.models.EEGGraphConvNet.rst	API doc — GCN model
docs/api/models/pyhealth.models.EEGGATConvNet.rst	API doc — GAT model
docs/api/tasks/pyhealth.tasks.EEGGCNNClassification.rst	API doc — classification task
docs/api/tasks/pyhealth.tasks.eeg_gcnn_disease_detection.rst	API doc — disease detection task

Test Plan

pytest pyhealth/tests/test_eeg_gcnn.py -v
# 23 tests, all passing, ~12 seconds, no real EEG data required