Authors:
Sergio A. Alvarez
1
and
Carolina Ruiz
2
Affiliations:
1
Boston College, United States
;
2
Worcester Polytechnic Institute, United States
Keyword(s):
Time Series, Clustering, Modeling, Markov, Data Mining, Sleep.
Related
Ontology
Subjects/Areas/Topics:
Applications and Services
;
Artificial Intelligence
;
Biomedical Engineering
;
Biomedical Signal Processing
;
Computer Vision, Visualization and Computer Graphics
;
Data Manipulation
;
Health Engineering and Technology Applications
;
Human-Computer Interaction
;
Medical Image Detection, Acquisition, Analysis and Processing
;
Methodologies and Methods
;
Neurocomputing
;
Neurotechnology, Electronics and Informatics
;
Pattern Recognition
;
Physiological Computing Systems
;
Physiological Processes and Bio-Signal Modeling, Non-Linear Dynamics
;
Sensor Networks
;
Soft Computing
Abstract:
This paper presents a new algorithm for time series dynamical modeling using probabilistic state-transition models, including Markov and semi-Markov chains and their variants with hidden states (HMM and HSMM). This algorithm is evaluated over a mixture of Markov sources, and is applied to the study of human sleep stage dynamics. The proposed technique iteratively groups data instances by dynamical similarity, while simultaneously inducing a state-transition model for each group. This simultaneous clustering and modeling approach reduces model variance by selectively pooling the data available for model induction according to dynamical characteristics. Our algorithm is thus well suited for applications such as sleep stage dynamics in which the number of transition events within each individual data instance is very small. The use of semi-Markov models within the proposed algorithm allows capturing non-exponential state durations that are observed in certain sleep stages. Preliminary r
esults obtained over a dataset of 875 human hypnograms are discussed.
(More)