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

In the field of brain‒computer interfaces (BCIs), developing a reliable machine learning (ML) model for real-time robotic hand control systems based on motor imagery (MI) brain signals requires substantial research. For this purpose, a set of ML models has been developed and tested to identify robust models via MI sensor data fusion under both nonadversarial and adversarial attack conditions. This paper addresses numerous essential areas, including the development of ML models for electroencephalography (EEG) MI signal datasets, with a focus on proper preprocessing and evaluation under both nonadversarial and adversarial attack conditions. Three phases make up the process. In the first phase, raw MI-EEG datasets from the Graz University BCI competition are identified and preprocessed. The preprocessing encompasses six key stages: EEG-MI signal filtering, segmentation, time‒frequency domain feature extraction, merging and labeling, normalization (resulting in Dataset I), and feature fusion (resulting in Dataset II). In the second phase, both datasets are used to develop nine different ML methods and are evaluated via nine performance metrics. These models are trained and tested against adversarial and nonadversarial scenarios. In the third phase, the fuzzy decision by opinion score method (FDOSM) and the multiperspective decision matrix (MPDM) are combined to benchmark the ML models via the fuzzy multicriteria decision-making (MCDM) approach. The random forest (RF) model achieved the best overall performance, with the lowest FDOSM scores: 0.18241 for Dataset I and 0.21636 for Dataset II. A lower FDOSM score means better results across all evaluation criteria. To further assess the developed methodology, the RF model was tested on Dataset III, comprising EEG data from four participants collected via the EMOTIV EPOC. The mean classification accuracy achieved by the RF model was 83 % with standard preprocessing, and it improved to 86 % with the application of feature fusion techniques. Additionally, this study employed the local interpretability model-agnostic explanation (LIME) method to provide an understanding of the RF model's behavior and enhance the interpretability of the results in the context of individual predictions.