Comparison of Methods for Hand Trajectory Estimation in 3D Space Using EEG Signals

This study investigates various methodologies for estimating hand trajectories from EEG signals, focusing on the comparative performance of different regression models and neural network architectures during a grasp-and-lift task. Using data from 9 subjects in the WAY-EEG-GAL dataset, which includes detailed EEG and movement data, we tested various methods to estimate hand trajectories. These included traditional techniques like Multivariate Linear Regression (MLR) and Partial Least Squares Regression (PLSR), as well as more advanced methods such as Multilayer Perceptrons (MLP), Convolutional Neural Networks (CNN), and a hybrid model combining CNN with Gated Recurrent Units (CNN-GRU). Our findings reveal that the CNN-GRU model provided the most accurate predictions of hand positions in 3D space, achieving a maximum Pearson Correlation Coefficient (PCC) of $0.87 \pm 0.02$ and a minimum Mean Squared Error (MSE) of $0.012 \pm 0.001$. Other neural networks, such as CNN and MLP, also significantly outperformed traditional methods like MLR and PLSR, though they require considerably higher computational resources. These results demonstrate the potential of using neural networks, particularly hybrid models like CNN-GRU, to effectively estimate hand trajectories based on EEG signals during motion execution. In the future, these methods could be adapted to predict imagined hand movements in individuals with disabilities, opening new possibilities for assistive technologies.