Finger joint reconstruction from electroencephalography signals with deep learning

Abstract

Brain-Computer Interface (BCI) is an important technique for robot control and rehabilitation. Motion Trajectory Prediction BCI (MTP-BCI) is considered suitable for high-precision tasks by decoding continuous motion information from Electroencephalography (EEG) signals. While numerous studies have focused on reconstructing the kinematics of upper and lower limbs, there is a dearth of updated research on finger movement reconstruction, particularly in non-invasive settings, despite its critical role in daily tasks. Utilizing deep learning neural networks, this study aims to reconstruct finger joint angles from EEG signals in a non-invasive setting. The models are rigorously evaluated on a novel dataset comprising 20 healthy subjects, and Explainable AI (XAI) algorithms are employed to interpret the parameters of the trained models. The findings demonstrate the feasibility of MTP-BCI for finger movement, potentially offering insights for asynchronous BCI and the development of new BCI commands. Additionally, this study is investigating the use of Motor Execution (ME) decoders on Motor Imagery (MI) data to decode imagery movement kinematics, based on neuroscience findings indicating overlaps between cortical regions related to motor imagery and execution. Future efforts will focus on: • assessing decoder robustness, particularly in rest state • enhancing ME decoding performance • designing a novel neural network architecture for continuous movement signal decoding This study is targeting to be submitted to the 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2024. Updates on progress, findings, and paper drafts will be continuously included into revised versions of this report.