Adaptive decoding for brain-machine interfaces through Bayesian parameter updates.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) translate neural activity into actuator movements.
• Neuronal representations drift over time, necessitating adaptive BMI decoders for sustained performance.
• Existing BMI decoders struggle with long-term accuracy due to neural plasticity and recording instability.

Purpose of the Study:

• To introduce and evaluate a novel Bayesian regression self-training method for adaptive BMI decoder updates.
• To assess the efficacy of this method in maintaining decoder accuracy without external movement information.
• To validate the approach through offline reconstructions and real-time closed-loop experiments.

Main Methods:

• Proposed a Bayesian regression self-training method to update unscented Kalman filter decoder parameters.
• Utilized joint and factorized formulations of Bayesian linear regression for neuronal tuning model updates.
• Evaluated performance using offline cursor control reconstructions and real-time closed-loop experiments in non-human primates.

Main Results:

• Bayesian regression self-training significantly improved offline reconstruction accuracy compared to non-updated decoders.
• Real-time closed-loop experiments demonstrated sustained BMI control accuracy over 29 days.
• The self-training method did not require hand movement data, desired movement assumptions, or goal information.

Conclusions:

• Bayesian regression self-training effectively adapts BMI decoders to neural signal variations.
• This adaptive approach ensures long-term BMI control accuracy, crucial for clinical neuroprosthetics.
• The method offers a robust solution for enhancing the reliability and applicability of brain-machine interfaces.