Continuous closed-loop decoder adaptation with a recursive maximum likelihood algorithm allows for rapid performance...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) require effective decoder adaptation for optimal performance.
• Current closed-loop decoder adaptation (CLDA) methods often rely on periodic updates, leading to delays in performance improvement and recalibration.
• Continuous adaptation offers a potential solution to enhance and maintain online BMI performance efficiently.

Purpose of the Study:

• To introduce and evaluate the Recursive Maximum Likelihood (RML) algorithm for continuous adaptation in Kalman filter decoders.
• To demonstrate the practical advantages and properties of RML for CLDA.
• To compare the performance of RML against existing CLDA algorithms in closed-loop experiments.

Main Methods:

• Developed the RML algorithm for continuous adaptation of Kalman filter parameters in BMIs.
• Implemented RML with a single, intuitive half-life parameter for real-time adaptation rate adjustment.
• Reformulated RML's recursive update rules for computational efficiency with large neural feature sets.
• Conducted closed-loop experiments with macaque monkeys performing a center-out reaching task using neural data (spiking activity or local field potentials) to control a 2D cursor.

Main Results:

• RML effectively leverages accurate batch-based updates while adapting parameters at every time step.
• The RML algorithm demonstrated efficient adaptation even with a large number of neural features due to its memory-efficient and computationally fast recursive rules.
• In closed-loop experiments, RML achieved higher performance levels more rapidly compared to a CLDA algorithm with intermediate timescale adaptation.
• Monkeys used either spiking activity or local field potentials to control a 2D cursor, demonstrating the algorithm's versatility.

Conclusions:

• RML is a highly effective CLDA algorithm for achieving rapid performance acquisition in BMIs.
• Continuous adaptation using RML offers significant advantages in reducing decoder training and recalibration times.
• The algorithm's intuitive parameterization and computational efficiency make it a practical choice for real-time BMI applications.