Mathematical programming
Extend optimal survival trees and survival support vector machines to competing risks, with objective functions and constraints that explicitly address fairness.
Current research project
Fair and interpretable survival models for kidney transplantation
Developing machine-learning methods that account for competing risks while improving predictive performance, transparency, and fairness in clinical transplant decisions.
The challenge
Better predictions must also be explainable and fair
Conventional survival models are valuable in health research but may be limited by linearity, proportional-hazards assumptions, and single-outcome formulations. In transplantation, competing events such as patient death and graft failure require models that represent clinical reality more faithfully.
Flexible machine-learning approaches can capture complex patterns, but their black-box behavior and susceptibility to bias can limit clinical adoption. This project addresses those tensions together rather than treating accuracy, interpretability, and fairness as separate objectives.
Core objective
Create survival models that support consequential decisions
The project develops and evaluates new survival-analysis methods for competing risks, combining mathematical programming and deep learning to improve discrimination, calibration, explainability, and fairness.
Research pathway
From model formulation to clinical utility
Two complementary modelling tracks converge in real-world decision simulations.
Deep survival learning
Develop architectures inspired by DeepSurv and DeepHit, using specialized learning strategies to represent competing outcomes and mitigate bias.
Explainability
Investigate how specific donor, recipient, and compatibility variables influence predictions through complementary explanation techniques.
Expected contribution
Evidence for responsible deployment
By evaluating predictive performance and downstream allocation outcomes together, the project aims to produce decision-support tools with demonstrable clinical utility and policy relevance - while reducing bias and making model behavior easier to understand.