Loading...

A New Coupled-HMM Framework with Applications in Multichannel Brain Signal Processing

Karimi, Sajjad | 2022

177 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 55915 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Shamsollahi, Mohammad Bagher
  7. Abstract:
  8. The human brain can be described as a dynamic system with multiple subsystems interacting with one another and multi-channel observations of these subsystems are available. The modeling of a system from its observations allows us to gain insight into how its various components interact with one another and also provides intuition about the desired system. Hidden Markov Model (HMM) is a probabilistic model with hidden states that is suitable for modeling these types of systems. Multi-channel observations are available from several subsystems interacting with each other within a general system. In this case, it may be necessary to develop more comprehensive models incorporating multi-channel structure and behavior. Coupled Hidden Markov Model (CHMM) is an appropriate model for multi-channel observations with spatio-temporal interactions between channels. The computational complexity and number of parameters for inference and learning problems in CHMM increase exponentially and rapidly as the number of channels increases. This research focuses on a particular type of CHMM model called the hidden structure influence model (LSIM), which does not suffer from the exponential growth of parameters. Despite this, existing algorithms for solving LSIM problems, such as inference and learning, are not accurate and lack theoretical and mathematical support. The current study proposes and develops a new framework for solving Markov problems in LSIM that can be applied and implemented in a large number of channels and limited samples of observations. The first proposed theory is related to the development of a new approximate inference algorithm with computational complexity instead of which has less error in approximations than existing approximation methods. In the second proposed theory, a new learning algorithm with convergence analysis and reestimation algorithms is presented, which has a strong theoretical and mathematical support similar to HMM Therefore, in the new framework, it is possible to implement LSIM for datasets with more than 100 channels, and the performance improvement of LSIM with the proposed framework is statistically significant compared to HMM and CHMM in conventional applications of modeling and classification based on the results. In the continuation of this research, two new applications for LSIM in the processing of multi-channel brain signals are presented. The first application of combining multi-channel information is in the problem of sleep stage detection. In this application, LSIM achieves 87.3% accuracy on Sleep-EDF data with integrated three-channel integration, which is one of the best results. The second application is related to the determination of effective brain connections, which is proposed using LSIM and Monte-Carlo method, a new estimator for multivariate transfer entropy. The simulation results of the neural mass model show that the proposed method increases AUCPR values by 3.7% on average in the range of zero to five dB
  9. Keywords:
  10. Hidden Markov Model ; Coupled Hidden Markov Model ; Brain Connectivity ; Multivariate Transfer Entropy ; Latent Structure Influence Model ; Multi-Channel Information Fusion ; Multi-Channel Brain Signals

 Digital Object List

 Bookmark

...see more