EEGaitPredict: predicting gait from high density pre-frontal EEG electrodes

Abstract

Background: Stroke-induced gait impairments are one of the leading causes of reduced mobility and functional independence among survivors, with many patients experiencing abnormal walking patterns such as hemiparetic gait and foot drop. While conventional rehabilitation approaches can partially restore locomotion, they often plateau after the initial recovery phase and are resource-intensive. This has prompted growing interest in EEG-based brain-computer interface (BCI) systems to decode motor intention directly from neural activity to assist gait rehabilitation. However, most existing BCI studies rely on full-head (FH) EEG, with limited investigation into more practical configurations such as prefrontal cortex (PFC)-only EEG despite evidence on its core involvement in gait planning Objective: The objective of this project is to investigate whether EEG signals collected exclusively from the prefrontal cortex can predict continuous lower-limb joint angles with accuracy comparable to full-head EEG. The study also explores how different walking conditions affect decoding performance, Method: A multimodal experimental framework was developed to capture synchronized EEG and joint kinematics from 16 healthy participants. EEG signals were recorded using both a full-head 64-channel cap and an 8-channel prefrontal EEG system, while continuous joint angles were obtained using wireless goniometers. Subjects performed a 10-meter walking task under four different conditions: Free, Mindful, Slow, and Fast walking. The collected EEG data were preprocessed and used to train a deep learning model (GaitNet) to predict joint angles. Performance was evaluated using Pearson’s correlation coefficient (r-value) Results: preliminary result points that the highest decoding performance is achieved during Free and Mindful walking conditions, with mean r-values of 0.59 and 0.591, respectively. Slow walking produced the lowest decoding performance (mean r = 0.417), while Fast walking exhibited higher variance but occasional high r-values (up to 0.802). These findings suggest that moderate-paced walking yields more consistent and decodable EEG patterns. Across all conditions, some trials exhibited poor r-values, highlighting variability in trial-level signal quality and decoding robustness. Conclusion: Preliminary findings indicate that decoding lower-limb kinematics from EEG is most effective during moderate, self-paced walking scenarios. However, these conclusions are based solely on full-head EEG data and intra-block validation. Further analysis—including inter-condition generalization and decoding using only PFC EEG—is still underway. These upcoming evaluations are necessary to determine whether PFC EEG can effectively substitute full-head EEG