CT-adaptTL

This is the PyTorch implementation of the Explainable Cross-Task Adaptive Transfer Learning for Motor Imagery EEG Classification.

Flowchart of pre-training, fine-tuning and explainability analysis

The aim of this work is to explore the feasibility and interpretability of cross-task knowledge transfer between MI and ME. This can largely relax the constraint of training samples for MI BCIs and thus has important practical sense.

Resources

Datasets

HGD: Link

openBMI: Link

GIST:Link

Sample pre-trained models

For openBMI: Link

For GIST: Link

Sample fine-tuned models

For openBMI: Link

For GIST: Link

Instructions

Install the dependencies

It is recommended to create a virtual environment with python version 3.6 and running the following:

pip install -r requirements.txt

You can choose a suitable download source to ensure the success of the download.

Obtain the raw dataset

Download the raw dataset from the resources above(Please download the ME/MI data in mat file format), and save them to the same folder.

    datasets/GIST/s01.mat
                 /s02.mat
                 /...

    datasets/openBMI/sess01_subj01_EEG_MI.mat
                    /sess01_subj02_EEG_MI.mat
                    /...
                    /sess02_subj01_EEG_MI.mat
                    /sess02_subj02_EEG_MI.mat
                    /...

    datasets/HGD/test
                    -/1.mat
                    -/2.mat
                    -/...
                /train
                    -/1.mat
                    -/2.mat
                    -/...

Data alignment

When HGD is the source domain and openBMI is the target domain,run

    PROCESS14_54.py
    
    process54.py

Please create folders DATA14_54 and DATA54 in the process directory to store the processed data

When HGD is the source domain and GIST is the target domain, run

    PROCESS14_52.py
    
    process52.py

Please create folders DATA14_52 and DATA52 in the process directory to store the processed data

Model pre-training

When openBMI is the target domain, run

    base14_54.py

When GIST is the target domain, run

    base14.py

This process is likely to take some time. We have provided sample pre-trained models in above resources

Subject-specfic

When openBMI is the target domain, run

    specific_54.py

When GIST is the target domain, run

    specific_52.py

Subject-independent

When openBMI is the target domain, run

    subject_independent14_54.py     

When GIST is the target domain, run

    subject_independent14_52.py

Adaptive fine-tuning

To fine-tune the pre-trained model with openBMI dataset, run:

    subject_adaptive14_54_sin.py

To fine-tune the pre-trained model with GIST dataset, run:

    subject_adaptive14_52_sin.py

This process is likely to take some time. We have provided sample fine-tuned models for each subject in above resources .

Model explaining

To explain the fine-tuned models with openBMI dateset, run:

    shap_value_adjust_54.py     

To explain the fine-tuned models with GIST dateset, run:

    shap_value_adjust.py

Results

The classification results for our method in three scenarios are as follows:

openBMI

Methodology	Mean (SD)	Median	Range (Max-Min)
Subject-Specific	73.48(16.16)	69.50	53.00(100.00-47.00)
Subject-Independent	69.00(15.58)	64.50	51.00(99.00-48.00)
Subject-Adaptive	76.59(15.93)	74.50	51.00(100.00-49.00)

GIST

Methodology	Mean (SD)	Median	Range (Max-Min)
Subject-Specific	61.48(12.84)	57.50	55.00(99.00-44.00)
Subject-Independent	58.96(10.39)	56.50	39.00(82.00-43.00)
Subject-Adaptive	69.77(12.54)	67.00	51.00(100.00-49.00)

Cite:

If used, please cite:

Miao M, Yang Z, Zeng H, Zhang W, Xu B, Hu W. Explainable cross-task adaptive transfer learning for motor imagery EEG classification. Journal of Neural Engineering. 2023. DOI 10.1088/1741-2552/ad0c61

Acknowledgment

We thank Kaishuo Zhang et al and Schirrmeister et al for their wonderful works.

Zhang, Kaishuo, et al. "Adaptive transfer learning for EEG motor imagery classification with deep convolutional neural network." Neural Networks 136 (2021): 1-10.https://doi.org/10.1016/j.neunet.2020.12.013

Schirrmeister, Robin Tibor, et al. "Deep learning with convolutional neural networks for EEG decoding and visualization." Human brain mapping 38.11 (2017): 5391-5420. https://doi.org/10.1002/hbm.23730