About Me

Hi, and welcome to my page! I’m a biomedical data scientist passionate about turning real-world clinical data into insights and tools that improve healthcare. I recently completed my PhD in Biomedical Informatics at Harvard Medical School, where I focused on developing, evaluating, and refining clinical tools for kidney disease.

My research spans population-level epidemiological studies using NHANES (N=44K) and predictive modeling with large scale longitudinal electronic health records (N=5M). I’ve led numerous research projects from conception through publication, resulting in several first-author papers, including a recent paper selected for oral presentation at the 2027 Pacific Symposium on Biocomputing. My technical expertise includes building end-to-end ML pipelines, from cohort extraction and feature engineering through model training and evaluation, and tackling challenges such as rare-event modeling and large-scale EHR modeling under memory and compute constraints. I’ve worked across data modalities and modeling approaches, from survival models using longitudinal laboratory measurements to neural networks applied to medical imaging.

As I wrap up my doctoral projects, I am actively seeking my next opportunity, where I hope to apply my expertise in data science and machine learning to real-world challenges in healthcare. (View Resume)

Outside of research, I maintain creative interests in music and ceramics.

Education

B.S. in Biomedical Engineering
Johns Hopkins University
Baltimore, MD

Education

B.M. in Vocal Performance
Peabody Conservatory
Baltimore, MD

Education

Ph.D. in Biomedical Informatics
Harvard Medical School
Boston, MA

Selected Research Projects

Laboratory Trajectories Improve Kidney Failure Risk Estimation

Current kidney failure risk equations rely on only a patient's most-recent laboratory values, even though patients often have years of routinely collected lab history. This project examines whether longitudinal lab trajectories can improve risk estimation while remaining interpretable and clinically grounded.

From a large EHR dataset spanning over 5 million individuals, we constructed a cohort of 270K patients with chronic kidney disease. We developed a trajectory-based extension of standard risk models that incorporates features derived from repeated laboratory measurements over time. These models improve identification of high-risk patients compared with latest-value approaches, particularly for longer-term prediction.

The results highlight how routinely collected laboratory histories contain meaningful signal beyond a single measurement, and how simple, interpretable trajectory features can improve risk stratification in clinical practice.
Read More

AI-clinician collaboration via disagreement prediction
A decision pipeline and retrospective analysis of real-world radiologist-AI interactions

Model development doesn’t end at deployment. In this project, which resulted in a first-author publication in Cell Reports Medicine, we explore how post-deployment data can be used to improve AI systems in practice.

This work develops a decision pipeline that integrates disagreement prediction with clinical significance and prediction quality to guide how AI outputs are used. Using real-world radiology data, we show that disagreement signals can be used to selectively escalate high-risk or uncertain cases, enabling a balance between reducing unnecessary clinician workload and maintaining diagnostic safety.
Read More

Quantifying the Clinical Implications of Imprecision in Kidney Function Estimates

Kidney function estimates guide diagnosis and treatment, yet a single number can obscure uncertainty that could change a clinical decision. This project examines the population-scale implications of that uncertainty and how it may affect confirmatory testing.

Using nationally representative NHANES data from 44,360 adults and published distributions of measured kidney function, we modeled how often estimated and measured values would produce different clinical classifications. At one commonly used threshold, we estimated discordant classifications for more than 14 million U.S. adults.

We also compared strategies for prioritizing confirmatory testing. Simple proximity to a threshold closely approximated probabilistic prioritization for kidney function thresholds, but rankings diverged for multivariable kidney failure risk estimates. These findings highlight the value of interpreting uncertainty in relation to the decision being made, helping identify when additional testing could change care.

Technical Areas

Machine learning on longitudinal healthcare data, survival analysis, rare-event prediction, and clinical impact evaluation

Preferred Programming Languages

  • Python (primary)
  • R

    • Tools

    • Cloud Computing (AWS, GCP)
    • Git
    • Docker
    • Deep Learning (PyTorch, Keras)
    To learn more about my journey and introduction to research, check out my story.