Application of Computer Vision and Explainability in EEG Classification for Motor Imagery Therapy after Stroke
JP García-Ortiz 1, Miroslav Zivanovic2, Marisol Gómez1,3
1NAIR Center, Pamplona, Spain
2 Public University of Navarra, Department of Electric Enginiering, Pamplona, Spain
3 Public University of Navarra, Department of Statistics, Informatics and mathematics, Pamplona, Spain
*Email: garcia.166837@e.unavarra.es
Introduction
Motor imagery (MI) has proven to be an effective strategy in the
rehabilitation of post-stroke patients by facilitating the activation
of preserved motor networks [1]. EEG signal classification using
artificial intelligence allows for the evaluation of mental exercise
adherence and effectiveness. However, current models still present
limitations. Gramian angular fields enable the application of
computer vision techniques to time series data [2]. In addition,
explainability methods such as Grad-CAM and Maximum Discrepancy
Distance (MDD) allow for interpretation of model decisions,
potentially optimizing electrode selection, improving the accuracy of
functional assessment, and providing new insights into cortical
activity during MI processes.
Methods
A public EEG dataset from 50 stroke patients was used, recorded with
32 electrodes during left- and right-hand MI tasks, with 20
repetitions of 2 seconds per class. The signals were preprocessed via
filtering to remove artifacts.
Subsequently, Gramian angular
fields were generated from each EEG segment. The resulting images
were input into a convolutional neural network (CNN) composed of 3
convolutional layers, 3 max pooling layers, one flattening layer, and
one final linear layer.
For model explainability analysis, the
MDD and Grad-CAM algorithms were applied. Evaluation metrics
included: accuracy, sensitivity, F1-score, and Chi-squared test.
Results
The model's ability to classify each patient's brain activity was
evaluated by training it individually with data from that patient.
MDD was used to identify the most relevant electrodes for
classification. A reduced dataset was generated using only images
from the most relevant electrodes, and the model was retrained to
assess classification performance again. The accuracy of the model
trained on the full dataset was compared to that of the model trained
only on images from the relevant electrodes. Finally, Grad-CAM was
used to analyze the most informative regions of the images for model
decision-making.
Discussion
Transforming EEG signals into Gramian angular fields enables the
application of advanced computer vision techniques for classification
and feature extraction, potentially enhancing post-stroke
rehabilitation approaches. The use of explainability algorithms such
as MDD allows for identification of key electrodes, optimizing
channel selection, simplifying analysis, and highlighting the most
informative signals. Furthermore, Grad-CAM contributes to model
interpretability by identifying critical image regions for
classification, improving physiological understanding and system
transparency.
Mishell Cadena Yánez, by her expertice in Explainable AI
1.- Braun, S., Beurskens, A., Kleynen, M., Schols, J., & Wade, D. (2008). Efficacy of motor imagery in post-stroke rehabilitation: a systematic review. Journal of NeuroEngineering and Rehabilitation, 5(8).
2º-Wang, Z., & Oates, T. (2015). Imaging time-series to improve classification and imputation. Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence, 3939-3945.
