Title: Bayesian sequential marginalization of a state space model for estimating motor unit numbers
Authors: Gareth Ridall - Lancaster University (United Kingdom) [presenting]
Abstract: This application comes from the field of neurology where an estimate the number of units supplying a muscle group is required. An increasing stimulus applied at the nerve each motor unit of the axon bundle is activated with increasing probability and the cumulative muscular response is recorded. A state space model with increasing dimension is used to model the increasing muscular response with stimulus. The observations are assumed Gaussian, conditional on the parameters and the latent binary firing indicators. In our state space model sufficient statistics and approximate sufficient statistics are used to model the state and measurement processes respectively. An efficient proposal system is used for the current firing indicators. These are re-weighted sequentially using the evidence as observations arrive. Lastly a residual resampling step is used to keep the numbers of possible weighted histories to a manageable number. Our methodology is substantially faster than Reversible jump Markov Chain Monte Carlo and is able to account for drift in the parameters for the measurement process.