Enhancing Cross-Session Motor Imagery Classification Using Graphical Models in ALS Patients

This repository contains the official implementation for our study:

"Enhancing Cross-Session Motor Imagery Classification Using Graphical Models in ALS Patients"
Rishan Patel, Barney Bryson, Tom Carlson, Dai Jiang, and Andreas Demosthenous
Under Review in tNSRE

---

🧠 Overview

This work addresses the challenge of non-stationarity in EEG-based Brain-Computer Interfaces (BCIs), particularly in longitudinal use cases for ALS patients. We explore the effectiveness of various graph-theoretic EEG representations; Phase Locking Value (PLV), Magnitude Squared Coherence (MSC), Cross Frequency Coupling (CFC), and Conditional Mutual Information (CMI) within a Graph Attention Network (GAT) framework to improve cross-session motor imagery classification.

---

📊 Key Results

• PLV-GAT outperforms conventional methods (CSP, EEGNet, DeepConvNet) in both ALS and healthy cohorts.
• Graph-based models exhibit enhanced robustness to session-to-session signal variability.
• Ablation studies identify subject-specific electrode configurations that further boost performance.
• PLV emerges as a computationally efficient, stable, and explainable candidate for real-time BCI.

---

🧬 Dataset

We evaluate our models on two datasets:

1. SHU Dataset — 25 healthy subjects, 5 sessions each, 32-channel EEG.
2. ALS Dataset — 8 ALS patients, 4 sessions each, 19-channel EEG, recorded using g.USBamp and BCI2000.

Details and access:

• ALS Dataset DOI: https://doi.org/10.5522/04/28156016.v1
• SHU Dataset: Referenced from [36] in the paper.

---

🧰 Requirements

• Python 3.8+
• PyTorch 2.x
• PyTorch Geometric 2.5.0
• NumPy, SciPy, Scikit-learn
• Matplotlib, Seaborn (for plots)