Nonlinear stochastic dynamic systems approach for personalized prognostics of cardiorespiratory disorders

Abstract

This research investigates an approach rooted in nonlinear stochastic dynamic systems principles for personalized prognostics of cardiorespiratory disorders in the emerging point-of-care (POC) treatment contexts. Such an approach necessitates new methods for (a) quantitative and personalized modeling of underlying cardiovascular system dynamics to serve as a virtual instrument to derive surrogate (hemodynamic) signals, (b) high-specificity diagnostics to identify and localize disorders, (c) real-time prediction to provide forecasts of impending disorder episodes, and (d) personalized prognosis of the short-term variations of the risk, necessary for effective treatment decisions, based on estimating the distribution of the times remaining till the onset of an anomaly episode. The specific contributions of the dissertation work are as follows:

1. Quantitative modeling for real-time synthesis of hemodynamic signals. Features extracted from ECG signals were used to construct atrioventricular excitation inputs to a nonlinear deterministic lumped parameter model of cardiovascular system dynamics. The model-derived hemodynamic signals, personalized to an individual's physiological and anatomical conditions, would lead to cost-effective virtual medical instruments necessary for personalized POC prognostics.

2. Random graph representation of the complex cardiac dynamics for disorder diagnostics. The quantifiers of a random walk on a network reconstructed from vectorcardiogram (VCG) were investigated for the detection and localization of cardiovascular disorders. Extensive tests with signals from PTB database of PhysioNet databank suggest that locations of myocardial infarction can be determined accurately (sensitivity of ~88% and specificity of ~92%) from tracking certain consistently estimated invariants of this random walk representation.

3. Nonparametric prediction modeling of disorder episodes. A Dirichlet process based mixture Gaussian process was utilized to track and forecast the evolution of the complex nonlinear and nonstationary cardiorespiratory dynamics underlying of the measured signal features and health states. Extensive sleep tests suggest that the method can predict an impending sleep apnea episode to accuracies (R^2) of 83% and 77% for 1 step and 3 step-ahead predictions, respectively.

4. Color-coded random graph representation of the state space for personalized prognostic modeling. The prognostic model used the stochastic evolution of the transition pathways from a normal state to an anomalous state in the color-coded state space network to estimate the distribution of the remaining useful life. The prognostic model was validated using the data from ECG Apnea Database (Physionet.org). The model can predict the estimated time till a disorder (apnea episode) onset to within 15% of the observed times 1-45 min ahead of their inception.