Overview
My research focuses on machine learning for electronic health records (EHR), with an emphasis on risk prediction for chronic and metabolic diseases such as MASLD. I am particularly interested in developing models that are both high-performing and clinically interpretable, while addressing issues of fairness, temporal structure, and real-world deployment in healthcare settings.
Project 1: Fairness-Aware Machine Learning for MASLD Prediction
I developed and evaluated multiple supervised machine learning models—including LASSO logistic regression, random forests, XGBoost, and neural networks—on a large-scale EHR cohort for MASLD prediction, achieving AUROC values up to 0.85.
To improve interpretability, I constructed a sparse LASSO-based model using the top 10 SHAP-selected features, maintaining strong predictive performance (AUROC = 0.84) while relying only on routinely collected primary care variables.
Additionally, I implemented fairness-aware postprocessing using equal opportunity constraints to reduce disparities in true positive rates across racial and ethnic subgroups, explicitly quantifying trade-offs between fairness and sensitivity.
Methods/tools: supervised machine learning, SHAP analysis, fairness evaluation (Fairlearn), PySpark, Python
Project 2 (In Progress): Temporal Graph Framework for Personalized Chronic Disease Risk Prediction
I am currently developing a temporal graph-based deep learning framework to model longitudinal patient trajectories from electronic health records (EHRs), with a focus on chronic disease risk prediction.
This work involves designing a personalized graph construction strategy that captures patient-specific clinical relationships and temporal dependencies across visits and medical events.
The model is being evaluated on chronic disease prediction tasks, with an emphasis on learning expressive patient-level temporal representations.
Methods/tools: graph neural networks (GNNs), graph convolutional networks (GCNs), temporal modeling, EHR reconstruction, Python
Project 3 (Future Work): Explainable Temporal Graph Neural Networks for Clinical Decision Support
Future work focuses on integrating explainability and interpretability techniques into temporal graph neural networks to improve transparency in patient-specific disease trajectory modeling.
The goal is to align learned temporal graph representations with clinically meaningful features, enabling more interpretable predictions and supporting clinical decision-making.
This direction aims to improve the trustworthiness and real-world usability of temporal graph models in healthcare environments.