Created by Albert Kjøller Jacobsen, Aron Djurhuus Jacobsen & Phillip Chavarria Højbjerg

This project investigates the role of data augmentation on the Temple University Hospital EEG Artifact corpus.

The outcome of the project is a student report that in detail describes our approach and discusses the outcome of our experiments.

The data considered is the open-sourced TUH EEG Artifact data set published by the Neural Engineering Data Consortium (NEDC).

This Git-repo contains python-files for pre-processing as well as fully implemented pipelines coded with an object-oriented approach for easily running Machine Learning fitting (using various data augmentation techniques). Furthermore, python scripts for obtaining results from pickled data files are to be found. Some parts of the code are redundant.

Code taken from David Enslev Nyrnbergs Git-repo, shared with us for this project. The preprocessing is done in the "prepData"-directory.

Original inspiration found in the benchmark paper on the TUH EEG Artifact data set published by Roy, S. (2019)

The models considered as well as the code for the augmentation methods can be seen in th "models"-directory. The hyperparameter-spaces considered for optimization of the models (using Bayesian Optimization) can be found in the "evaluation.pipeline.py"-file.

SMOTE / balancing pipeline as well as the augmentation-pipeline can be found in the "evaluation"-directory. Some of the files, namely the activeLearning_pipeline is currenly broken as it does not store the correct results and is neither optimized for the EEG approach. Active Learning pipeline includes coding parts taken from exercises in the course 02463 on Active Learning and agency.

Our results are saved as numpy-pickles using a function from the "prepData.dataLoader"-script. All results are found using a random seed set to 0 and can be reconstructed by reading the "paper/article" written for this project.