Trust and explainability in robotic hand control via adversarial multiple machine learning models with EEG sensor...

Area of Science:

• Neuroscience
• Machine Learning
• Robotics

Background:

• Developing reliable machine learning (ML) models for real-time robotic hand control using motor imagery (MI) brain signals is crucial for brain-computer interfaces (BCIs).
• Evaluating model robustness under both non-adversarial and adversarial conditions is essential for practical BCI applications.

Purpose of the Study:

• To develop and evaluate ML models for electroencephalography (EEG) MI signal datasets.
• To identify robust ML models for robotic hand control under adversarial attacks.
• To enhance the interpretability of ML model predictions in BCI systems.

Main Methods:

• Preprocessing of MI-EEG datasets including filtering, segmentation, time-frequency feature extraction, normalization, and feature fusion.
• Development and evaluation of nine ML models using nine performance metrics against adversarial and non-adversarial scenarios.
• Benchmarking ML models using fuzzy multicriteria decision-making (MCDM) approach with fuzzy decision by opinion score method (FDOSM) and multiperspective decision matrix (MPDM).

Main Results:

• The Random Forest (RF) model demonstrated the best overall performance, achieving the lowest FDOSM scores.
• Feature fusion techniques improved the RF model's classification accuracy from 83% to 86% on a separate EEG dataset.
• The Local Interpretability Model-agnostic Explanation (LIME) method was used to enhance the understanding of RF model predictions.

Conclusions:

• The developed ML models, particularly Random Forest, show promise for reliable robotic hand control in BCIs.
• Feature fusion is a valuable technique for improving the accuracy of EEG-based BCI systems.
• Model interpretability methods like LIME are important for understanding and trusting BCI system outputs.