Identifying multiscale hidden states to decode behavior.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-machine interfaces (BMIs) require accurate encoding models to translate neural activity into intended movement.
• Existing models often focus on single scales of neural activity (spikes or LFPs) or lack dynamic state representations.
• Recent findings suggest neural dynamics, not just direct representations, encode behavior, necessitating advanced models.

Purpose of the Study:

• To develop and validate an unsupervised learning algorithm for a multiscale dynamical encoding model with hidden states.
• To enable simultaneous characterization and decoding of discrete spikes and continuous field activity.
• To enhance BMI decoder performance by integrating information across neural scales.

Main Methods:

• Developed an unsupervised learning algorithm to estimate a multiscale state-space model with hidden states.
• Validated the model using simultaneous spike-local field potential (LFP) recordings during reaching movements.
• Utilized the learned model and decoder to identify hidden states and decode movement trajectories.

Main Results:

• The identified hidden states accurately decoded movement trajectories.
• Incorporating LFP data alongside spike data significantly improved decoding accuracy.
• Demonstrated that the unsupervised algorithm effectively integrates information across different neural scales.

Conclusions:

• The proposed unsupervised learning algorithm can effectively model multiscale neural activity and decode movement.
• Integrating spike and LFP data enhances decoding performance in BMIs.
• This algorithm offers a novel tool for studying neural encoding across scales and advancing BMI technology.