Domain-guided Self-supervision of EEG Data Improves Downstream Classification Performance and Generalizability

Authors: Neeraj Wagh, Jionghao Wei, Samarth Rawal, Brent Berry, Leland Barnard, Benjamin Brinkmann, Gregory Worrell, David Jones, Yogatheesan Varatharajah

Affiliation: University of Illinois at Urbana-Champaign, Mayo Clinic

Work accepted in proceedings of the ML4H Symposium 2021 with an oral spotlight!

• Paper: https://proceedings.mlr.press/v158/wagh21a.html
• ML4H Poster: https://drive.google.com/file/d/1-B7joEquzr4kqsfSGDqksUARWsxlXOyA/view?usp=sharing
• ML4H 5-minute Video: https://recorder-v3.slideslive.com/?share=56322&s=8950fb88-3f3f-4490-92ff-760b73e5a4b5
• ML4H Slides: https://drive.google.com/file/d/1uUma2e4GKIn8Zo7dEQRl_WkyVcBLzpaR/view?usp=sharing
• Final Models, Pre-computed Features, Training Metadata: Box
• Raw Data: MPI LEMON (no registration needed), TUH EEG Abnormal Corpus (needs registration)

Installation

• A dockerfile is provided with all Python dependencies.

Mapping naming conventions between the paper and code

• Models:
  • FULLSSL in code refers to "HS+BSE+AC" in the paper.
  • DBRTriplet refers to "BSE+AC".
  • BTTriplet refers to "HS+AC".
  • BTDBR refers to "HS+BSE".
  • Triplet refers to "AC only".
  • DBR refers to "BSE only".
  • BT refers to "HS only".
  • SOTA refers to "ShallowNet".
• Tasks:
  • "EEG Grade" is the "condition" task with TUH dataset.
  • "Eye State" is the "condition" task with Lemon dataset.

Command Line Arguments

• Input choices for tasks: "condition", "gender", "age".
• Input choices for dataset: "lemon", "tuh".
• Input choices for ablation models: "FULLSSL", "DBRTriplet", "BTTriplet", "BTDBR", "Triplet", "DBR", "BT".

How to Reproduce the Results Reported in the Paper

1. Download resources folder, which contains pre-computed feature arrays(power spectral density, topographical maps data, and preprocessed timeseries), metadata from Box. Place it in the root directory of the project.
2. Enter evaluation folder.
3. To evaluate linear baseline, run the following command:

python linear_baseline_eval.py --dataset={dataset choice} --task={task choice}

4. To evaluate proposed ablation models, run the following command:

python ablation_models_eval.py --gpu_idx={gpu index} --dataset={dataset choice} --task={task choice} --mode={ablation model choice}

5. To evaluate SOTA model, run the following command:

python SOTA_eval.py --gpu_idx={gpu index} --dataset={dataset choice} --task={task choice} 

How to Fine-tune Existing Pre-trained Models for Downstream Tasks

1. Download resources folder from Box. Place it in the root directory of the project.
2. Enter Fine-tune folder.
3. To fine-tune ablation models, run the following command:

python ablation_pipeline.py --gpu_idx={gpu index} --dataset={dataset choice} --task={task choice} --mode={ablation model choice}

4. To fine-tune SOTA model, run the following command:

python SOTA_pipeline.py --gpu_idx={gpu index} --dataset={dataset choice} --task={task choice}

How to Perform Supervised Learning for Downstream Tasks

1. Download resources folder from Box. Place it in the root directory of the project.
2. Enter supervised_learning folder.
3. To train the supervised learning models, run the following command:

python linear_baseline_train.py --dataset={dataset choice} --task={task choice}

[Work in Progress] - How to Perform Pre-training using Domain-guided Self-superivsed Tasks

[Work in Progress] - How to Use Your Own Dataset to Train the Self-supervised Learning Tasks

Contact

• Issues regarding non-reproducibility of results or support with the codebase should be emailed to Neeraj and John.
• Neeraj: nwagh2@illinois.edu / Website / Twitter / Google Scholar
• John: wei33@illinois.edu
• Yoga: varatha2@illinois.edu / Website / Google Scholar

Citation

Wagh, N., Wei, J., Rawal, S., Berry, B., Barnard, L., Brinkmann, B., Worrell, G., Jones, D. & Varatharajah, Y.. (2021). Domain-guided Self-supervision of EEG Data Improves Downstream Classification Performance and Generalizability. Proceedings of Machine Learning for Health, in Proceedings of Machine Learning Research 158:130-142 Available from https://proceedings.mlr.press/v158/wagh21a.html.