Decoding lower-limb kinematic parameters during pedaling tasks using deep learning approaches and EEG.

Area of Science:

• Neuroscience
• Robotics
• Biomedical Engineering

Background:

• Stroke impairs mobility, necessitating advanced rehabilitation tools like brain-computer interfaces (BCIs).
• Restoring gait function using BCIs with robotic systems, such as motorized mini exercise bikes (MMEBs), shows promise.
• Accurate kinematic estimation from electroencephalography (EEG) signals for continuous motion remains a significant challenge.

Purpose of the Study:

• To compare two artificial neural network (ANN) decoders for estimating lower-limb kinematic parameters during pedaling.
• To evaluate the feasibility of using deep learning (DL) for continuous decoding in BCI-controlled MMEBs.

Main Methods:

• Utilized Long Short-Term Memory (LSTM), a type of recurrent neural network (RNN), to decode EEG features.
• Estimated ankle position (x, y) and knee joint angle during pedaling tasks.
• Analyzed kinematic variance and correlations between pedaling speed and decoder performance.

Main Results:

• LSTM achieved a Pearson correlation coefficient (PCC) of approximately 0.58 for kinematic parameter reconstruction from delta-band EEG features.
• The proposed algorithm demonstrated effectiveness in distinguishing pedaling and rest periods.
• A negative linear correlation was observed between pedaling speed and decoder performance, suggesting easier estimation at slower speeds.

Conclusions:

• Deep learning methods are feasible for estimating lower-limb kinematics from EEG signals during pedaling.
• This research facilitates the development of more robust MMEB controllers for BCIs using continuous decoding.
• Findings support enhanced, personalized rehabilitation by maximizing degrees of freedom in BCI-driven systems.