EEG decoding and analysis of reach-to-grasping using an explainable artificial intelligence approach

Motor decoding from electroencephalography (EEG) using deep learning models has gained significant attention in applications like brain-computer interfaces. While these models offer high accuracy, their ”black-box” nature limits their transparency, which is especially problematic in neurophysiological contexts where understanding the decision-making process of the these models is essential. To address this challenge, this thesis focuses on the analysis and decoding of EEG signals during a reach-to-grasping task using an explainable artificial intelligence approach. The study involved the acquisition and processing of EEG signals from 19 healthy participants, in which participants performed movements to reach and grasp four different objects. Two classification tasks were considered: movement vs rest and the discrimination between three types of grasping (power, intermediate, and precision). After the EEG preprocessing, movement-related cortical potentials were extracted and analyzed to assess whether EEG signals are modulated in the conditions under analysis. EEGNet, a convolutional neural network widely used for EEG signal classification, was then used and its performance evaluated. Five post-hoc explainability methods were then applied to understand the spatio-temporal features that EEGNet identifies as relevant for decoding motor activity of the first task (movement vs rest). These were: saliency maps, deepLIFT, input × gradient, gradient shap and integrated gradients. The outputs of the explainability methods were compared with movement-related cortical potentials dynamics, which are known to reflect the neurophysiological correlates of movement. DeepLIFT, input × gradient, gradient shap, and integrated gradients emerged as the most reliable explainers for this analysis. This thesis provides a preliminary analysis on post-hoc explanation techniques for improving the comprehension of deep learning models applied to motor EEG decoders.