Five active externally funded projects as PI or MPI — three NIH R01s and two Eli Lilly industry awards totaling $9.7M — plus co-investigator roles on NIH RC2 and U54 awards.
Active projects · PI / MPI
1R01LM015078 · 2026 – 2031 · Total award $3,497,380
We are developing, validating, and prototyping DischargeWatch, a system that applies multimodal machine learning to structured EHRs, unstructured clinical notes, and audit logs capturing team activities to enable accurate 24-hour discharge predictions. A performance-variation assessment and optimization framework integrating value-sensitive design and transfer learning addresses performance differences across sociodemographic groups, and a human-centered approach — silent model testing in a simulation environment, interest-holder advisory panels, and iterative prototyping — will deliver a high-fidelity clinical decision support prototype ready for future clinical trials.
1R01LM015131 · 2026 – 2030 · Total award $1,521,723
We are establishing a human-AI collaboration framework that integrates large language models, graph neural networks, role-specific AI agents, and human-in-the-loop methods to systematically identify, analyze, and address inefficiencies in healthcare professional interaction networks. We first fine-tune LLMs to extract clinically meaningful EHR tasks and map interactions among healthcare professionals, then apply GNN-based analyses to detect and explain inefficiencies — including the specific nodes, edges, and subnetworks correlated with patient outcomes — culminating in iterative collaboration between AI agents and interest-holder advisory panels to translate data-driven insights into actionable criteria. The project will deliver an informatics framework of fine-tuned LLMs, trained GNNs, AI agents, and HITL protocols that turns complex EHR data into actionable guidance, helping healthcare professionals optimize care coordination and ultimately improve patient outcomes and care quality.
1R01LM014199 (PIs: Li, Chen, Su) · 2023 – 2028 · Total award $3,118,680
This project develops novel machine-learning methodologies and robust evidence to study drug/gene/adverse drug event associations. Using BioVU resources, we compare patients who experienced adverse reactions from drug-drug interactions with those who did not, focusing on the genetic differences that make some patients more susceptible — including heightened risks of cardiac and renal complications. The resulting pharmacogenetic findings provide a valuable resource for prospective studies and contribute toward precision medicine and improved clinical care.
Lilly Grant Office A-39787 · 2025 – 2027 · Total award $1,000,000
Leveraging longitudinal EHR data, we distinguish patients receiving continuous obesity care from those lacking it, analyze audit logs to determine which providers perform key clinical activities, and employ human-AI collaboration for root-cause analysis — combining role-specific AI agents representing patient, clinician, and health-system perspectives with expert panel reviews — before designing and implementing feasible interventions targeting the identified barriers. Our CareContinuum iOS app grew out of this line of research.
Lilly Grant Office A-42361 · 2026 – 2028 · Total award $599,757
We are refining an AI event-and-timestamp extraction pipeline to quantify operational gaps in early Alzheimer's care — from biomarker confirmation to first infusion — identifying root causes through quantitative workflow metrics and stakeholder input, and deploying a triage bot, an MRI safety agent, and a generative prior-authorization agent within routine VUMC workflows to directly target the bottlenecks. Read the VUMC Reporter story.
Active projects · Co-Investigator
Creating the Urinary Stone Disease Hub with de-identified data from over 230,000 individuals with kidney stone disease across nine U.S. health systems.
Ensuring AI/ML in the Bridge2AI program is developed and applied in an ethical and trustworthy manner through an iterative Scaffold–Assess–Facilitate–Evaluate cycle.
Defining the relationship between ICU delirium and dementia to pave the way for preventive programs, strategic rehabilitation, and targeted interventions for ICU survivors.
A multi-ethnic, multi-omic, longitudinal study of how gut microbiota and their metabolites affect cardiometabolic outcomes after metabolic and bariatric surgery.
Completed
Used machine learning to infer care coordination patterns and their variability in relation to hospital length of stay, then engaged healthcare experts through interviews and surveys to assess which data-driven solutions health systems could adopt.
Discovered effective disease-specific care teams, learned dependencies between providers for resource allocation, and modeled treatment workflows to find the most efficient sequences of care — reducing length of stay and costs.