Effective prediction – as opposed to detection – of future states of a complex system remains a challenge, mainly due to diverse combinations of nonlinear and nonstationary characteristics exhibited by the underlying system. In particular, the challenge is caused by the nondeterministic or stochastic behaviors of the processes, the complex relationships among measured signals and the underlying system states, and the irregular evolution of the system dynamics. To tackle that challenge, Dr. Le’s research uses a stochastic nonlinear dynamic systems approach to (i) characterize the system’s dynamics and its state transitions via the development of a multivariate state space reconstructed from the measured signal features, and (ii) develop a personalized prognostic approach to providing a statistical distribution, in real time, of the onset of an impending episode based on a nonparametric statistical Dirichlet-Process Mixture-Gaussian-Process (DPMG) model. The timeliness and effectiveness of the intervention can be enhanced substantially if an impending acute episode can be predicted before the clinical symptoms become evident. 

The research involves monitoring the evolutionary mechanism of the electrodynamics in a physiological system using multidimensional bio-signal harnessed from sensors, therefrom forecasting the occurrences of the acute disorder onsets before the clinical symptoms appear. It uses CCAST’s HPC systems for data collection, storage, and processing.

The proposed framework consists of three parts: (i) disorder characterization [1, 2], (ii) feature representation [1, 3], and (iii) disorder onset forecasting [3–6]. Features characterized for the progression are chosen using iterative feature extraction processes with the criteria selected by machine-learning based classification performance. They are then transformed into a dynamical graph where each node represents a state variable and the edges denote the evolution of these points in the state space which represents the disorder progression over the time. Such representation yields cross-crossing transition and unbalancing of overpopulation and sparsity nodes in the state space due to the intrinsic complex dynamics of the disorders. The Laplacian-eigen-projection based method is employed to account for the scattering and overcrowding of the adjacent nodes in the state space. Here, the state vectors are projected to account for the significant feature variation in the feature space. Finally, a Generalized Dirichlet Mixture Gaussian Prediction method is used to cluster all the similar statistical distribution of disorder state variable and forecast the evolution of each state in the defined state space. The estimation of the transition from normal states to an anomalous state is utilized to calculate the distribution of the remaining time until the onset of an impending abnormal episode. 

The above methods have been validated in various case studies of characterizing and forecasting the onset of obstructive sleep apnea, atrial fibrillation and epileptic seizures from large electronic clinical health records and online databases such as PhysioNet.