PhD Thesis Defense: Rumeng Li, AI for Aging: Longitudinal Disease Risk Modeling from Clinical Data
Content
Speaker:
Abstract:
Population aging is increasing the burden of aging-associated chronic diseases, many of which develop gradually over years before formal diagnosis. Characterizing risk earlier in the disease course may support more timely clinical evaluation, monitoring, and intervention planning. Electronic health records (EHRs) offer a longitudinal view of disease progression through routine clinical care. However, many early signals are fragmented across years of unstructured clinical narratives, making them difficult to extract, integrate, and interpret at scale. This dissertation develops artificial intelligence (AI) and clinical natural language processing (NLP) methods to transform these longitudinal narratives into scalable and interpretable models of disease risk.
Alzheimer disease (AD) serves as the primary application because of its growing public health burden and prolonged prediagnostic period, during which cognitive, behavioral, functional, and social changes may emerge years before diagnosis. Using large-scale real-world data from the Veterans Health Administration (VHA) and Mass General Brigham (MGB), this work characterizes longitudinal AD-related signs and symptoms documented in clinical narratives and evaluates their predictive value beyond structured EHR features.
Extracting these disease-related signals at scale requires high-quality labeled clinical data, which remain difficult to obtain because access to clinical data is restricted by privacy and regulatory requirements and expert annotation is costly. To address this bottleneck, this dissertation develops LLM-based approaches for synthetic data generation that leverage clinical knowledge, population characteristics, and longitudinal symptom patterns to expand limited labeled datasets and improve the extraction of disease-related signals from clinical narratives. The extracted signals are then integrated across clinical domains and time through a multi-agent framework for longitudinal AD risk assessment.
The dissertation further examines social and behavioral context and cross-system generalizability. Longitudinal health-related social and behavioral signals provide complementary predictive information beyond conventional clinical variables, while cross-system analyses evaluate the consistency of narrative signals and the transportability of risk models across healthcare systems with different patient populations and documentation practices.
Together, this dissertation establishes a framework for modeling disease risk from longitudinal clinical narratives, spanning disease-specific signal extraction, LLM-based synthetic data generation, multi-agent risk assessment, incorporation of longitudinal social and behavioral context, and evaluation across healthcare systems. Although developed primarily in the context of AD and aging, these methods provide a foundation for studying longitudinal risk in other chronic diseases.
Advisor:
Hong Yu