PROJECTS

PH.D. RESEARCH

Toward Generalizable Hospital Length-of-Stay Prediction Using Hierarchical Modeling with Uncertainty Quantification and Continuous Learning

Overview 

Hospital Length-of-Stay (LOS) is a critical healthcare metric. It drives resource allocation, patient outcomes, operational efficiency, and care quality. Yet prediction models trained at one hospital routinely fail when deployed at another, because they ignore the nested structure of healthcare data where patients nested within hospitals, hospitals nested within regions.

This dissertation develops a framework that models this hierarchical structure explicitly, delivering predictions that generalize across diverse hospitals without site-specific retraining, paired with reliable uncertainty bounds and continuous adaptation as healthcare evolves.

Aim 1: Quantifying the Hierarchical Structure of LOS Variation 

This study uses multilevel modeling to clarify how much hospital‑level factors and regional characteristics each contribute to patients’ length of stay. By separating variation across hospitals from broader geographic effects, the research shows that both organizational practices and local context play important roles in how long patients remain hospitalized. The results highlight the need for policies that address not only individual hospitals but also regional health system differences. 

Read the Paper — IISE Transactions on Healthcare Systems Engineering 

Aim 2: Identifying the Optimal Hierarchical Modeling Architecture  

This research systematically compares three major approaches for modeling hierarchical data: traditional statistical models, tree‑based methods, and neural networks. It evaluates how well each approach captures multilevel structure, handles variability across groups, and balances accuracy with interpretability. The study highlights the strengths and limitations of each class of models and provides guidance on choosing appropriate techniques for complex, nested datasets. 

Read the paper — Machine Learning with Applications 

Aim 3: Robust Uncertainty Quantification for Hierarchical Predictions  

This study introduces a hybrid uncertainty‑aware framework for predicting hospital length of stay using hierarchical random forests. By combining Bayesian uncertainty estimates with conformal prediction, the method produces reliable and adaptive prediction intervals that work across thousands of hospitals. The approach achieves near‑nominal coverage, adapts interval width based on case difficulty, and generalizes well to new hospital settings, supporting safer and more efficient clinical decision‑making. 

Read the paper — Scientific Reports 

Aim 4: Continuous Model Adaptation Across Time and New Hospitals 

This work introduces an adaptive framework for hospital length‑of‑stay prediction that updates over time and across sites without full retraining. The mixed‑effects incremental gradient boosting model combines population‑level learning with hospital‑specific adjustments and adaptive conformal prediction intervals. Evaluated on over three million discharges from hundreds of hospitals, the approach matches full retraining accuracy while cutting computation nearly in half and improving stability under temporal drift, especially for hospitals with the largest prediction challenges. 

Under Review

M.S. RESEARCH

Developing Artificial Intelligence tools to investigate the phenotypes and correlates of chronic kidney disease patients in West Virginia 

My thesis focuses on building AI‑based tools to study chronic kidney disease in West Virginia, a region with high CKD burden. Using statewide clinical data, I develop models that identify patient phenotypes and examine the social, behavioral, and medical factors associated with disease outcomes. The work provides insights into local health disparities and supports data‑driven strategies for improving CKD care in the state. 

Read thesis — West Virginia University

Read the paper— Healthcare Analytics

COLLABORATIVE RESEARCH

Geometry-Aware Uncertainty Quantification via Conformal Prediction on Manifolds 

Developed adaptive geodesic conformal prediction for responses on Riemannian manifolds. The method replaces Euclidean residuals with geodesic nonconformity scores and adapts to local prediction difficulty. Applied to geomagnetic field forecasting using IGRF-14 satellite data. 

[Read the paper — ICML] 

COLLABORATIVE RESEARCH

Hierarchical Conformal Prediction with Finite-Sample Coverage  

Extended conformal prediction to hierarchical data structures, achieving finite-sample coverage guarantees with improved efficiency across scales. 

[Read the paper — Results in Applied Mathematics]