Real-time particle filtering and smoothing algorithms for detecting abrupt changes in neural ensemble spike activity

Area of Science:

• Neuroscience
• Computational Neuroscience
• Time Series Analysis

Background:

• Sequential change-point detection is crucial for neuroscience applications like seizure and pain detection.
• Previous methods using Poisson linear dynamical systems were limited by Gaussian noise assumptions.
• Online brain-machine interface (BMI) applications require efficient recursive filtering for latent variable tracking.

Purpose of the Study:

• To improve the speed and accuracy of change-point detection in neuroscience.
• To introduce non-Gaussian dynamical noise and accommodate non-Gaussian likelihoods for enhanced modeling.
• To develop efficient particle filtering and smoothing algorithms for real-time applications.

Main Methods:

• Proposed particle filtering and smoothing algorithms to estimate the state posterior with non-Gaussian noise and likelihood.
• Implemented algorithms using graphics processing unit (GPU) computing for accelerated computation.
• Validated algorithms through computer simulations and experimental data for acute pain detection.

Main Results:

• Novel sequential Monte Carlo methods enable rapid detection of stochastic jump processes in population spike activity.
• The new approach demonstrates robustness against spike sorting noise and varying signal-to-noise ratios.
• GPU implementation facilitates parallel processing for real-time computational performance.

Conclusions:

• The developed algorithms offer a significant advancement in sequential change-point detection for closed-loop neuroscience.
• The methods are suitable for real-time BMI applications, enhancing detection speed and reliability.
• This work provides a robust framework for analyzing neuronal ensemble activity and detecting critical events.