Making brain-machine interfaces robust to future neural variability.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Clinical translation of brain-machine interfaces (BMIs) is hindered by decoders that fail when neural recording conditions change.
• Current decoders often require frequent retraining due to neural variability.

Purpose of the Study:

• To develop a more robust BMI decoder resistant to future neural variability.
• To improve the reliability and reduce downtime of BMIs in clinical applications.

Main Methods:

• A novel multiplicative recurrent neural network (RNN) BMI decoder was developed.
• The decoder was trained on extensive historical data and synthetic perturbations to handle diverse recording conditions.
• Performance was evaluated using a non-human primate preclinical BMI model.

Main Results:

• The new RNN decoder learned various neural-to-kinematic mappings effectively.
• Training with larger and more diverse datasets significantly improved decoder robustness.
• The developed decoder demonstrated robustness in conditions that rendered a state-of-the-art Kalman filter decoder ineffective.

Conclusions:

• Accumulated data history can be effectively harnessed to create robust BMI decoders.
• This new BMI strategy may significantly reduce decoder retraining downtime, facilitating reliable BMI use.
• The findings support a promising new direction for advancing BMI clinical translation.