Projects

Kidney transplantation and exchange

A collaboration with Dr. Stanislaw Stepkowski and Dr. Dulat Bekbolsynov at the University of Toledo Medical Center, running since 2017. We use national registry data from the Scientific Registry of Transplant Recipients (SRTR) together with molecular HLA matching to ask a practical question: which donor-recipient pairs will actually last, and how should organs be allocated so that more patients, and more underserved patients, get a lasting graft?

• Immunogenicity-aware matching. Scoring donor-recipient compatibility at the eplet and physiochemical level instead of the antigen level, and showing that low-immunogenicity donors improve long-term survival, including for African American recipients who have historically been disadvantaged by antigen-level matching.
• Allocation geometry. Evaluating a 250 nautical mile concentric circle allocation system and the matching algorithm that goes with it, to improve the pool of compatible donors a center can draw from.
• From model to clinic. Dan Schellhas's dissertation (Ph.D. 2026) closes the gap between data-driven eplet mismatch models and what a transplant clinician can interpret and trust. Desmond Offei's dissertation (Ph.D. 2024) asked whether eplet load and electrostatic mismatch scores matter in practice, using machine learning on registry outcomes.

This work has produced journal papers in Frontiers in Immunology, Computers in Biology and Medicine, Transplant Immunology, and Transplantology, plus abstracts at the American Transplant Congress. See Research for the full list.

Data-driven public safety

A collaboration with BGSU's Center for Justice Research, where we bring forecasting, mapping, and data engineering to questions criminologists are already asking.

• Crime forecasting by tract. Time-series models (Prophet) fit per census tract and crime category for Toledo and Dayton, producing forecasts and interactive reports that let analysts see where trends are moving rather than where they have been.
• An interactive map of crime, calls, and demographics. A Shiny application that overlays crime and call-for-service data on census tracts for Toledo, with demographic change over time, counts versus rates per thousand residents, and police sector boundaries.

Generative models for small and imbalanced medical data

Clinical datasets are small, imbalanced, and expensive to label. Iman Khazrak's doctoral work asks how far generative models can go toward fixing that, and what it costs to adapt them.

• Diffusion models as data augmentation. Synthetic vocal fold pathology images from a DDPM improved classification in a pilot study published in the European Archives of Oto-Rhino-Laryngology, and a broader comparison of generative models for small, imbalanced medical image sets appeared in Artificial Intelligence and Applications.
• Parameter-efficient fine-tuning of diffusion models. Two papers at ICAI 2026: one on the rank trade-offs when fine-tuning diffusion models with LoRA, and LoRA-Diffusion, which decomposes the low-rank update trajectory itself.
• Domain adaptation. Adaptive gradient-norm weighting for domain adversarial training, in the International Journal of Data Science and Analytics, so models trained at one site transfer to another.
• AI-augmented query optimization. A survey and taxonomy, under review at ACM Computing Surveys, of how learned components are changing database query optimizers.

Working with me

Students who join the lab work on one of the threads above. Expect version control from day one, weekly meetings, papers written in LaTeX, and a project that is meant to ship to a real user, not just a committee. If one of these fits, email me at least three semesters before you plan to graduate.