Research
Research Philosophy
My work is grounded in pragmatic innovation. I develop and evaluate applied artificial intelligence methods with an emphasis on predictive performance, fairness, uncertainty, and reliability under real-world conditions. My goal is to bridge methodological advances in AI with meaningful clinical application by asking not only whether a model performs well, but when its predictions are sufficiently trustworthy to inform patient care, for whom they work well, and where they fail.
Across this work, I am particularly interested in methods that make model limitations visible rather than obscuring them behind aggregate performance metrics. Ultimately, my research aims to support clinical AI systems whose benefits extend equitably across patient populations while remaining useful and interpretable in practice.
Current Research & Applied Work
Data Science Intern | UPHS Data Science, PennDnA at Penn Medicine | 2026–Present
- Contribute to translational and applied AI, machine learning, and data science initiatives, including the development and evaluation of models, systems, and software, with consideration of opportunity, feasibility, and governance.
Doctoral Research
PhD, Information Science | Drexel University | 2026
- Advisor: Christopher C. Yang, PhD
- Committee: Quyen Ngo, PhD; Judy Wawira Gichoya, MD, MS; Bhupesh Shetty, PhD; Erjia Yan, PhD; Grace Lu-Yao, PhD
- Dissertation: A Framework for Fair and Robust Clinical Risk Prediction through Collaborative Learning and Localized Uncertainty Quantification (DOI)
My doctoral research focused on the development of methods for fair and reliable clinical prediction. This work examined how subgroup fairness can be improved without substantially sacrificing predictive utility and how uncertainty can be characterized at the level of individual predictions rather than treated only as a population-level property.
Two central methodological contributions were a collaborative learning framework that treats patient subgroups as distinct learning clients during model optimization and a localized conformal prediction framework that adapts uncertainty estimates to the difficulty of individual cases.
Funding & Fellowships
NIH AIM-AHEAD Research Fellowship (2022)
Supporting research in health equity, algorithmic fairness, and trustworthy AI.Edith Peterson Mitchell, MD Health Equity Travel Scholarship (2023, 2024)
ECOG-ACRIN Cancer Research Group Meetings.
Core Focus Areas
Clinical AI & Predictive Modeling
Developing and evaluating predictive models for clinical decision support using real-world health data, with emphasis on clinically meaningful performance and translation to practice.
Fairness, Robustness & Uncertainty
Developing methods to characterize and improve subgroup fairness, model robustness, calibration, and uncertainty in high-stakes clinical AI systems, including approaches that account for variation in prediction difficulty across individual patients.
Foundation Models in Healthcare
Evaluating emerging foundation models, including large language models and tabular foundation models, for clinical applications, with emphasis on reliability, reasoning, calibration, fairness, reproducibility, and real-world readiness.
Real-World Evidence & Health Disparities
Using large-scale EHR, registry, and administrative data to characterize healthcare disparities, evaluate patterns of care, and generate evidence about differences in treatment and outcomes across patient populations.
Selected Projects
Reliable and Robust Clinical AI
Algorithmic Fairness & Uncertainty Quantification
- Designed fairness auditing frameworks for clinical risk prediction models across chronic kidney disease, substance use disorder, and oncology datasets, integrating novel mitigation techniques and benchmarking against existing approaches (JCO CCI, AIME 2025, IEEE ICHI 2024, IEEE ICHI 2023).
- Leveraged over 10 years of clinical nursing experience to guide feature engineering and problem formulation, distinguishing true physiological signal from artifacts of clinical workflow.
- Quantified disparities in readmission prediction and treatment completion, demonstrating how commonly used fairness metrics fail to capture systemic healthcare inequities.
- Developed a Neighborhood-Adaptive Difficulty Score combining k-nearest neighbor topology with conformal prediction to distinguish model limitations from inherent case complexity.
(Manuscript in preparation)
Tabular Foundation Models for Clinical Prediction
Performance, Fairness, Calibration & Uncertainty
- Evaluating tabular foundation models, including TabPFN, TabICL, and TabFM, against established machine learning approaches on real-world clinical prediction tasks, with emphasis on predictive performance, calibration, subgroup fairness, and uncertainty quantification.
- Investigating the use of conformal prediction and group-specific calibration to characterize uncertainty and subgroup reliability, including whether foundation models provide more efficient and equitable prediction sets than conventional models.
- Initial evaluation on MIMIC-IV sepsis mortality prediction found tabular foundation models to be competitive with or outperform tuned classical baselines across multiple dimensions of clinical model performance and reliability.
(Poster presented at the Emory HITI Lab Symposium 2026; manuscript in preparation)
Clinical NLP & Large Language Model Evaluation
- Developed EnsReas, an iterative prompting framework that improved medical question-answering accuracy and consistency on USMLE datasets (+3-5%) across both closed (GPT-4) and open-source clinical LLMs (JAMIA).
- Applied open-source clinical LLMs to extract pathologic TNM cancer stage from real-world pathology reports without labeled training data, demonstrating that prompting and ensemble-based reasoning can achieve competitive performance and improved consistency relative to fine-tuned BERT baselines (AIME 2024, IEEE ICHI 2024).
- Designed and evaluated an LLM-assisted pipeline to automate extraction of population demographics from biomedical literature, enabling scalable equity analysis (IJDC, IDCC 2026 poster 1, IDCC 2026 poster 2).
Real-World Evidence & Health Disparities Modeling
- Developed a mixed-order Markov chain simulation using VA Corporate Data Warehouse data to model prostate cancer treatment sequences and identify empirically observed pathways associated with elevated mortality risk across racial groups.
(Manuscripts in preparation) - Conducted a national registry study using the AAO IRIS® Registry to quantify racial and gender disparities in retinal vein occlusion treatment via multivariable logistic regression, highlighting inequities in access to anti-VEGF therapy (Ophthalmology Retina).