Decoding micro-electrocorticographic signals by using explainable 3D convolutional neural network to predict finger...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Traditional methods for decoding finger movements from electroencephalography (EEG) and electrocorticography (ECoG) relied on manual feature extraction from bandpass power changes.
• Existing machine learning models for movement decoding often lack interpretability, functioning as 'black boxes'.

Purpose of the Study:

• To introduce a novel 3D convolutional neural network (3D-CNN) for decoding finger movements using ECoG data.
• To enhance the interpretability of brain-computer interface (BCI) models using explainable AI (xAI) techniques.
• To identify the physiological relevance of specific brain signal features in motor control.

Main Methods:

• Developed a 3D-CNN model trained on ECoG data from epilepsy patients undergoing awake craniotomy.
• Processed ECoG signals to extract power spectral density across multiple frequency bands, forming a 3D data matrix.
• Integrated adaptive explainable AI (xAI) techniques, including Grad-CAM and SHAP, for model interpretation.

Main Results:

• The 3D-CNN model achieved high accuracy in predicting finger movements, with root-mean-square errors (RMSE) ranging from 0.20-0.38.
• Explainable AI identified the high gamma (HG) band as critical for movement prediction.
• Specific cortical regions involved in distinct finger movements were elucidated through xAI analysis, confirming the HG band's significance in motor control.

Conclusions:

• The 3D-CNN model combined with xAI significantly improves finger movement decoding accuracy from ECoG data.
• This approach offers an efficient and interpretable solution for brain-computer interface (BCI) applications.
• The study reinforces the crucial role of the high gamma (HG) band in motor control.