My research can be loosely grouped into Kidney Exchange and AI Ethics.
Patients with kidney failure have only two options: a lifetime on dialysis, or kidney transplantation. Dialysis is far more expensive and burdensome than transplantation, however donor kidneys are scarce — on average, 20 people die each day in the US while waiting for a transplant. Furthermore, many patients in need of a kidney have willing living donors, but cannot undergo transplantation due to medical incompatibilities.
To address this supply-demand mismatch, kidney exchange allows patients with willing living donors to swap their donors in order to find a compatible (or better) patient-donor match. Formulated as an optimization problem, kidney exchange is NP-hard and APX-hard, though modern exchanges are solvable in a reasonable amount of time (due to efficient formulations such as PICEF and PC-TSP).
In addition to being computationally hard, kidney exchange raises several logistical and ethical challenges. My research has focused on these challenges, particularly in fairness for marginalized patients, robustness to uncertainty, and moral challenges of designing a kidney exchange mechanism.
Fairness in Kidney Exchange
How can we prioritize marginalized patients, without severely impacting the overall exchange? This is the question we address in our AAAI 2018 paper. We study previous methods of enforcing fairness (and their unintended consequences), propose a new method that avoids undesirable behavior, and demonstrate the effects with data from fielded kidney exchanges.
- Duncan C. McElfresh, and John P. Dickerson, "Balancing lexicographic fairness and a utilitarian objective with application to kidney exchange." Conference on Artificial Intelligence (AAAI), 2018. (link)
Robustness in Kidney Exchange
There are many sources of uncertainty in real kidney exchanges — due to medical, moral, and policy factors. We considers two types of uncertainty in kidney exchange: edge-weight uncertainty
(i.e. uncertainty in transplant quality), and edge-existence uncertainty
(i.e. whether or not a transplant can occur). We propose efficient robust-optimization approaches to both types of uncertainty, and characterize their performance on real and synthetic data.
- Duncan C. McElfresh, Hoda Bidkhori, and John P. Dickerson, Scalable Robust Kidney Exchange.” Conference on Artificial Intelligence (AAAI), 2019.
Ethics and Kidney Exchange
Designing a kidney exchange program requires input from medical professionals, policymakers, computer scientists, and ethicists. A "good" kidney exchange program should be both technically- and morally-sound — however these objectives can sometimes be at odds with each other, and communication between a diverse group of experts can be challenging. Our working paper (under review) proposes a division of labor for the kidney design process, outlining responsibilities for experts of different disciplines. We apply the framework of Li et al.
to our context, analyze existing mechanisms, and propose future work from both an ethical and technical perspective.
- Duncan C. McElfresh, Vincent Conitzer, and John P. Dickerson. “Ethics and Mechanism Design in Kidney Exchange.” (Working paper).
- Duncan C. McElfresh, Patricia Mayer, Gabriel Schnickel, and John P. Dickerson. "Ok Google: Who Gets the Kidney?": Artificial Intelligence and Transplant Algorithms. Panel presentation and discussion at the annual meeting of the American Society of Bioethics and Humanities (ASBH), Anaheim, CA. (Forthcoming)
Algorithms increasingly inform — and sometimes make — important decisions with moral implications. Popular conceptions of AI-ethics dilemmas often involve futuristic systems, or idealized scenarios like the trolley problem. I'm interested in ethical implications of real AI systems that are currently used to make consequential decisions, in fields such as kidney exchange , criminal sentencing, and hiring. Below I outline some of my ongoing projects.
Ethical Algorithm Design
Designing an algorithm with potentially ethical implications is always difficult, especially when algorithm designers (who often aren't ethicists) and stakeholders (who often aren't computer scientists) must work together to build a system that is both technically- and ethically-sound. To this end, we are building a framework for exploring the technical and moral landscape — for a variety of applications, such as designing kidney exchange policy. This work is inspired by the framework of Li et al.
AI and End-of-Life Care
Patients who are faced with end-of-life care decisions (such as whether or not to receive life-sustaining treatment) are sometimes unable to communicate their wishes to their doctor. In this case, if the patient has not previously recorded these wishes (i.e. in an advance directive), then decision making is left to their surrogate decision maker — often a close family member. However surrogates often don't know what the patient wants, and can make decisions that go against a patient's wishes. We are using AI to better understand patient preferences for end-of-life care — to make better care decisions, and to reduce the burden on surrogates.