Identifying Evolving Multivariate Dynamics in Individual and Cohort Time Series...

SHARE:

Event Speaker: 

Shamim Nemati

Event Location: 

Grier B, 34-401B

Event Date/Time: 

Tuesday, December 11, 2012 - 4:00pm

 
Identifying Evolving Multivariate Dynamics in Individual and Cohort Time Series, with Application to
Physiological Control Systems
 
Physiological control systems involve multiple interacting variables operating in feedback loops that enhance an organism’s ability to self-regulate and respond to internal and external disturbances. The resulting multivariate time-series often exhibit rich dynamical patterns, which are altered under pathological conditions. However, model identification for physiological systems is complicated by measurement artifacts and changes between operating regimes. The overall aim of this thesis is to develop and validate computational tools for identification and analysis of structured multivariate models of physiological dynamics in individual and cohort time-series.
 
We first address the identification and stability of the respiratory chemoreflex system, which is key to the pathogenesis of sleep-induced periodic breathing and apnea. Using data from both an animal model of periodic breathing, as well as human recordings from clinical sleep studies, we demonstrate that model-based analysis of the interactions involved in spontaneous breathing can characterize the dynamics of the respiratory control system, and provide a useful tool for quantifying the contribution of various dynamic factors to ventilatory instability. The techniques have suggested novel approaches to titration of combination-therapies, and clinical evaluations are now underway.
 
We then study shared multivariate dynamics in physiological cohort time-series, assuming that the time-series are generated by switching among a finite collection of physiologically constrained dynamical models. Patients whose time-series exhibit similar dynamics may be grouped for monitoring and outcome prediction. We develop a novel parallelizable machine learning algorithm for outcome-discriminative identification of the switching dynamics, using a probabilistic dynamic Bayesian network to initialize a deterministic neural network classifier.
 
In validation studies involving simulated data and human laboratory recordings, the new technique significantly outperforms the standard Expectation Maximization approach for identification of switching dynamics. In a clinical application, we show the prognostic value of assessing evolving dynamics in blood pressure time-series to predict mortality in a cohort of intensive care unit patients.
 
A better understanding of the dynamics of physiological systems in both health and disease may enable clinicians to direct therapeutic interventions targeted to specific underlying mechanisms. The techniques developed in this thesis are general, and can be extended to other domains involving multi-dimensional cohort time-series.
 
Thesis Supervisors:
George Verghese and Atul Malhotra