LiverMatch · Applied Machine Learning for Liver Transplantation

Clinical collaboration

LiverMatch was one of three Georgia Tech student teams collaborating with Dr. Marwan Kazimi of Emory University School of Medicine on the broader liver-transplantation problem. Weekly meetings provided clinical context and feedback on project direction and results. Our team focused on donor-liver characterization, recipient and transplant-related data analysis, and exploratory graft-failure risk modeling.

Problem

Liver-offer acceptance varies substantially across transplant centers. Repeated declines can delay allocation, allow viable organs to expire, or cause them to reach recipients later than intended. The team narrowed that broad allocation problem to two tractable questions: how to characterize donor-liver profiles and how donor, recipient, and transplant variables relate to graft-failure outcomes.

Data

The project used OPTN/SRTR datasets covering deceased donors and liver transplant or waitlist records. The deceased-donor dataset contained 308,308 records and 539 features. Preparation required merging data files with separate label files, normalizing several forms of missing or unknown values, selecting clinically relevant features, removing unusable rows, and one-hot encoding categorical variables.

Methods

The team explored:

• K-means clustering and Gaussian mixture models for donor characterization
• A feed-forward neural network for predicting cluster assignments
• Random forest regression for single-value graft-failure-time prediction
• Random survival forest modeling for time-varying graft-failure risk
• Kaplan–Meier analysis for exploratory survival comparisons

Project progression

1. Proposal: Problem definition and planned methodology.
2. Midterm checkpoint: Dataset preparation, subproblem decomposition, and initial modeling.
3. Final project: Refined donor characterization and exploratory graft-failure modeling.

Results

The project selected a Gaussian mixture model over K-means for donor characterization. A neural classifier was used to test whether the learned group assignments were recoverable. K-means produced near-perfect classification behavior, but its loss did not converge, so the team considered the result unreliable rather than evidence of a superior model.

Random forest regression overfit and did not perform well. The random survival forest reported an approximate concordance index of 0.916. That figure is exploratory: only 419 records remained after preprocessing, and one label comparison contained 411 samples in one class and eight in the other. Kaplan–Meier analysis was used to inspect survival behavior across risk groups, not to establish clinical validity.

Limitations

• The work was exploratory and the post-processing graft-failure sample was small.
• Class imbalance limited confidence in parts of the evaluation.
• The models were not deployed clinically or used to make transplant decisions.
• The project did not produce a completed liver-allocation optimizer.
• The reported metrics should not be interpreted as scientific or clinical validation.

My role

This was a team project. My documented contributions were dataset preparation and analysis, task formulation, model and hyperparameter debugging, the final report and presentation, and project website development. The broader modeling and analysis described above reflects the team's combined work, not work I completed independently.

Artifacts

• Proposal (PDF) — Problem definition and planned methodology.
• Midterm checkpoint (PDF) — Dataset preparation, subproblem decomposition, and initial modeling.
• Final project (PDF) — Refined donor characterization and exploratory graft-failure modeling.

External links

• Original project page
• Dr. Marwan Kazimi’s Emory profile