Research Thrust 1: Optimal Integrative Strategies for Prediction and Management of Multiple Chronic Conditions

Treatment for people living with Multiple Chronic Conditions (MCC) currently accounts for an estimated 66 percent of the Nation's healthcare costs and will continue to grow. While traditional epidemiological approaches have led to important findings of disease links and comorbidity associations, they are limited in their ability to effectively characterize patterns of MCC emergence over time and identify the variables that affect changes in the trajectory of health status. Using a massive retrospective healthcare dataset, this line of research investigates big data analytics and high-performance computing along with clinical expertise to understand the process of MCC emergence and progression over me.

  • We develop large scale Sparse Latent Regression Hidden Markov Clustering (SLR-HMC) algorithms to identify major patterns of comorbidity in MCC patients. To estimate the unknown parameters of SLR-HMC, we extend the classical Baum–Welch algorithm with Markov Clustering, and proximal gradient descent algorithm.

  • We investigate efficient structure learning algorithms for hierarchical temporal Bayesian network (HTBN) to reveal patterns of interaction among MCC and patient-level risk factors. We extend the longest path algorithm (LPA) to identify the most likely sequence of the comorbidity emerging from, or leading to, the evolution of a new chronic condition.

  • We explore Continuous Time Bayesian Network (CTBN) with flexible conditional dependencies to uncover the complex stochastic process governing the evolution of comorbidities in MCC patients with respect to their risk factors.

Research grants: This area of research has so far resulted in a $441,000 grant from National Institutes of Health (1SC2GM118266-01), a $66,234 grant from Department of Veteran Affairs (VA268-15-D-0073), and a $20,000 Research Advancement and Transformation grant from University of Texas at San Antonio (UTSA).

 

The unsupervised HTBN of conditional dependencies among OUD, HTBN, TBI, PTSD, and depression, along with the most likely path to the evolution of OUD (red path)

Snapshots of our graph summarization algorithm for pruning conditional dependencies among OUD, HTBN, TBI, PTSD, and depression

 

Research Thrust 2: Active Learning Methodology for Design and Optimization of Complex Expensive Tests

In an increasing number of cases involving estimation of a response surface, one is often confronted with situations where there are several factors to be evaluated but experiments are prohibitively expensive. In such scenarios, learning algorithms can actively query the user or other resources to determine the most informative settings to be tested. This area of my research aims at exploring active learning methodologies to minimize the data required for design and optimization of expensive tests for statistical estimation of complex systems.

  • We investigate the fundamental idea of adding a similarity learning penalty to design of experiments to simultaneously shrink the weight of less significant factor, while looking for most informative settings to be tested. We have also developed simple sequential design strategies for efficient determination of subsequent experiments based on the information from previous experiments. To leverage the intrinsic geometry of factor settings in highly nonlinear spaces, we adapt the methodology to nonlinear and black-box problems using gradient directed kernels.

 

Research grants: This research line has so far resulted in $371,937 Young Investigator Award of Air force Office of Scientific Research (AFOSR-YIP) (FA9550-16-1-0171). It has also brought about a grant of $99,723 from Air Force Research Laboratory (FA8650-13-C-5800).

 

 

Illustration of CTBN for 5 conditions including Substance Abuse based on the preliminary analysis: The thickness of the edges represents the strength of the conditional intensities.

GP   

SLRD

Kriging

KBAR, Variant 1

KBAR, Variant 2

 

Research Thrust 3: Image and Sensor-Driven Modeling of Complex Degrading Systems

Multi-stream sensor and image data can be effectively used for monitoring, diagnostics and prognostics of complex degrading systems in discrete and continuous manufacturing.

  • We investigate image-based process monitoring for speedy detection of various types of defects in high-speed discrete manufacturing systems. To leverage the local geometry of defects in spatial domain, we develop area-based spatio-temporal monitoring schemes based on Area Delaunay Triangulation of prediction error.

  • we develop UAV-based sensor and image technology for real-time monitoring of the health condition of structures and pipeline over extreme distances and rugged terrain.

  • We work on scalable multi-sensor prognostic degradation methodology that utilizes advanced dimensionality reduction of massive multi-stream degradation signals to predict remaining life distributions of complex manufacturing system.

Research grants: This research thrust has resulted in several industry grants, including three grants of $77,630, $77,000 and $90,000 from Harland Clarke Company, and a $84,272 grant from Flat Rock Engineering and Environmental LTD.

Triangulated area surface of an in control image

Triangulated area surface of a defective image