The progression of many chronic diseases, such as Type 1 diabetes (T1D), manifests dynamic processes that can be modified by environmental exposures. Modeling of the underlying disease progression holds critical values for better understanding of disease development, effective monitoring, and prevention. While the emerging big data studying these diseases provide great resources, current pace for translating these data into effective monitoring and intervention strategies has been slow due to the analytic challenges caused by potential multi-layer characteristics of disease progression processes, the high-dimensional exogenous factors, heterogeneous biomarker signals, and the complexity of continuous-time stochastic processes. To mitigate these challenges with the development of new models and computational algorithms, this research will provide the desired personalized monitoring and risk factor identification capability, which is crucial not only for increasing the situational awareness of the individuals who are at risk, but also for providing evidences for design, validation, and deployment of intervention strategies. Its generic nature will also help effective monitoring of many other dynamic systems in engineering and life sciences. The interdisciplinary nature of this research across data-driven risk monitoring, dynamic systems, high-dimensional variable selection, and healthcare, will prepare students with a diversified education background.
The objective of this project is to create a generic suite of computational approaches that can be applied for modeling, learning, and monitoring a set of dynamic diseases, whose progression processes may be modified by exogenous factors such as environmental exposures. Several methodological contributions are expected, including: (1) a novel rule-based monitoring methodology to convert high-dimensional complex biomarkers into disease risk evaluation, via the development of an efficient screening method for high-throughput rule discovery and an optimal design method for risk monitoring; (2) a multi-layer dynamic model that can investigate how the exogenous risk factors regulate the disease process, with integration of sparse multi-task learning to mitigate the high-dimensionality of exogenous factors; and (3) a high-dimensional robust risk factor identification framework that can identify exogenous factors with integration of knowledge learned from historical data, new measurements, and clinician's prognostics. These proposed methods will be evaluated with a practical example studying T1D in partnership with The Environmental Determinant of Diabetes in the Young (TEDDY) study.
