Evaluating Computational Brain Models as Dimensionality Reduction Methods for EEG

Author: Lorenz Heiler
Supervisors: Dr. Pedro Mediano, Dr. Gregory Scott
Institution: Imperial College London (MSc Individual Project)

Project Overview

This thesis benchmarks four families of Computational Brain Models (CBMs)—cortico-thalamic, Jansen-Rit, Wong-Wang, and Hopf—as dimensionality reduction tools for clinical EEG.

These mechanistic models are compared against traditional data-driven baselines, including PCA, spectral autoencoders, EEGNet-style autoencoders, and the catch22 feature set.

Key Contributions

• Unified Benchmark: A modular pipeline built on the Temple University Hospital Abnormal EEG Corpus (TUH-AB).
• Hybrid Approach: Implementation of amortized parameter-inference for the cortico-thalamic model, achieving 78.4% accuracy in abnormality screening while maintaining physiological interpretability.
• Comparative Analysis: Evaluation of latent space quality based on dimensionality efficiency, geometry preservation, and information content.

Repository Structure

• /code: Core implementation, model training, and parameter inference.
  • See the Detailed Code README for execution instructions.
• /Datasets: EEG data and preprocessing artifacts.
• /model_comparison_results_complete: Logs and performance metrics for all evaluated models.
• /testing: Validation scripts and unit tests.

Dataset

This project utilizes the TUH Abnormal EEG Corpus (v3.0.1), consisting of 2,993 sessions.