University of Texas Arlington

Patients with lung tumors generally cannot bear a breath-holding process in PET/CAT Scan. They have to breathe normally during a scan and an expensive way has to be applied to process the PET/CT imaging. To minimize respiratory motion artifacts an efficient quiescent period gating (QPG) method was developed at UW medical center to extract PET/CT data from the end-expiration quiescent period. However, it has been found that some portion of subjects cannot benefit from the QPG method. In this project, we studied the respiratory pattern effects on the PET/CT imaging quality, made extensive statistical pattern analysis and multivariate regression analysis, and finally constructed an important clinical recommendation system to discriminate the benefit and non-benefit subjects for the QPG scan approach. The paper on this work has been submitted to Physics in Medicine and Biology in June 2013. 

This research is to develop new mathematical models to predict events from nonstationary multivariate time series data. Time series data accounts for a large fraction of the world’s supply of data, and have very broad applications in engineering, economy, industrial manufactory, finance, management and many other fields. However, most of the current time series methods make an assumption of stationarity. They are unable to identify complex temporal patterns from nonstationary time series that are nonperiodic, nonlinear, irregular, and chaotic. Thus this research is motivated to develop new efficient approaches to analyze nonstationary time series patterns and predict critical events. Three original and fundamental contributions have been made in the research.
1. A novel piecewise-linear approximation algorithm for time series data using a data-driven decomposition strategy. The algorithm is capable of obtaining a piecewise-linear model for any arbitrary time series (stationary or nonstationary) efficiently and robustly at a very low computational cost. This work has been submitted to Pattern Recognition Letters.
2. An efficient online monitoring and segmentation framework for time series temporal patterns. The algorithm is an extension of the piecewise-linear approximation algorithm. It is capable of processing and segmentation of online time series streams in time O(1) using a set of closed-form mathematical formulas.
3. An adaptive learning framework for online pattern discovery and event prediction in multivariate time series data. The framework integrates time series segmentation, temporal pattern extraction, adaptive learning, and probability theory into an online pattern discovery and prediction system. This is a general pattern discovery framework, and has been evaluated on two challenging EEG prediction problems: epileptic seizure prediction and mental-state prediction. The proposed approach generated the most promising prediction results in both problems compared the current existing approaches.

The link between mental states/cognitive activities and brainwave recordings is of paramount importance with practical scientific and clinical implications. However, EEG-based brainwave patterns are extremely difficult to decode due to the nonstationary and chaotic properties. We conducted three related research projects on this problem category. In the first project, we successfully developed a data-mining framework to analyze EEG patterns and identify the mental state prior to erroneous keystrokes during a human typing experiment in critical tasks. This work has been published in IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans. In the second project, an adaptive pattern-learning framework was constructed to detect the mental states from drivers in a simulated driving environment. This work is to be submitted to IEEE Transactions on Intelligent Transportation Systems in Aug. 2013. The third project is a Google-funded project. We constructed an information retrieval system to identify cognitive and memory workload in different levels of task engagement using an EEG-based brain-computer interface.

Due to the great inter-individual variability, there has been a desperate need for personalized seizure prediction for patients with epilepsy using Electroencephalogram (EEG) recordings. However, it has been an unsolved challenging problem for long since it is extremely difficult to capture predictive patterns for each individual from chaotic multi-channel EEG time series. I made an important methodology breakthrough in this area by constructing an online adaptive pattern-learning framework. The proposed new approach combines feature extraction, feature selection, pattern modeling and identification, adaptive online learning, and feedback control theory to achieve personalized pattern learning and prediction. This study is among the pioneer studies to investigate adaptive learning mechanisms for epileptic seizure prediction, and has achieved very promising prediction performance on 10 patients with epilepsy. This work has been accepted for publication in IEEE Transactions on Knowledge and Data Engineering. This work also won the Finalist of INFORMS 2012 Data Mining Best Student Paper Award