PosthocXAI4BCI

Code for the paper "Towards Optimising EEG Decoding using Post-hoc Explanations and Domain Knowledge", EMBC 2024

Installing dependencies

Use requirements.txt with Python 3.7 or higher

Reproducing results

Dataset: To skip the pre-processing step to generate epoch data, you may refer to this repo for the dataset.

To use Conformer architecture with 3 conditions:

1. Train the model using all channel data
2. Using MI relevant data
3. Using feature relevance refer to the file Conformer_top_16_subs.ipynb. Change the class EEGMMIDTrSet and EEGMMIDTsSet to ensure the correct choice of channels.

extractResults.ipynb helps extract the model performance and save the results in .csv format

Notebooks GradCAM_MIchannels.ipynb helps generate feature relevance explanations in the form of topomaps and TF plots as visualised in the paper.