Time-frequency exploration of non stationary physiological signals: In search of meaningful distinctive traits.

摘要

This discussion is concerned with the analysis, interpretation and understanding of physiological signals. When plotted as a function of time, a physiological signal is likely to be in the form of a fluctuating trace with short-term and long-term structures, seemingly of never ending nature, nearly repetitive, but never quite so. Such signals can be processed and 'understood' in terms of statistical distributions for some purposes or in terms of deterministic chaotic dynamical systems for other purposes.; There exists a large body of previous work on the predictive continuation of time series of single observables. This present work is concerned with understanding what happens in the multivariate case, when there are interactions between variables and also for circumstances where those are no inter-variable interactions. In the latter case, we have a set of independent single observables. We illustrate and demonstrate these matters using computer synthesized data and also publicly available benchmark data. We used neural networks for implementation of the computational procedures. We describe one way of dealing with multivariate systems with interactions between variables. We use a system of local nets to ensure that system states addressed in retrieval are of the same nature as those used in the formulation of the predictor. Predicted continuations obtained in that manner are demonstrably of higher accuracy than those built without consideration of inter-variable interactions.; We also looked into the matter of identifying distinctive traits of physiological data sources in other ways. The Empirical Mode Decomposition (EMD) approach was evaluated using a body of data for a patient with sleep disorder. In that approach, time series data are reformulated and extracted in the form of Intrinsic Mode Functions (IMF). For respiration signals, joint time-frequency spectra for some specific Intrinsic Modes were particularly distinctive and effective in indicating differences between normal, transitional and Apnea stages of sleep. Such data and analysis would be of considerable value for monitoring and predicting those various different stages of sleep. That mode-decomposition approach as implemented with our methodology is promising and worthy of further exploration.