Pain Classification on EEG and Tabular Data

University project at the Lab for AI in Medicine at TU Munich with Felix Hsieh, Carlos Llano and Ivan Iliash (me). Winter Semester 22/23.

Project structure

├── docs          # Documentation files
├── framework     # Training framework code
├── notebooks     # Notebooks for exploratory data analysis
└── README.md

Data

EEG recordings and Tabular data (demographic information, disease relevant information) from healthy subjects and patients with chronic pain. The follwing data files were provided by the University Hospital rechts der Isar. Unfortunately the datasets can't be shared with the public.

data_Dinh_fs200.pt
data_Heitmann_fs200.pt
labels_Dinh_fs200.pt
labels_Heitmann_fs200.pt
tabular_Dinh.xlsx
tabular_Heitmann.xlsx

The EEG data is provided as batches of torch tensors in the .pt files.

For more details, see notebook EDA.ipynb.

Prerequisites

Create a python environment and install dependencies.
With virtualenv (https://virtualenvwrapper.readthedocs.io/en/latest/)

mkvirtualenv eeg-multimodal
pip install -r requirements.txt

With conda

conda env create -f environment.yml

Run training

Configure training changing values in framework/runner/config.json. Follow convention in config-structure.json.

Execute:

python framework/runner/runner.py
python framework/runner/runner_crossval.py

Training results will be saved in a folder named training_results/

Conclusions

Theta and Gamma are important frequencies bands for pain classification.
Tabular data use shows no improvements, which requires further exploration with complete datasets or additional features.
High confidence for Fibromyalgia classification caused by clear patterns in the EEG data (See strong oscillations).
Signal data augmentations applied to EEG can help when dealing with a small dataset.

For more details, see poster.pdf.