Mukund Telukunta, a doctoral candidate in computer science, will defend their dissertation titled “Eliciting and Learning Fairness Preferences of AI Models in Kidney Transplantation.” Their advisor, Dr. Venkata Sriram Siddhardh Nadendla, is an assistant professor in the computer science department.  The dissertation abstract is provided below.

Modern kidney transplantation incorporates artificial intelligence (AI) decision-support systems which exhibit social discrimination due to biases inherited from training data. Although researchers have proposed various group-based fairness notions to assess biases in AI, it remains uncertain which criterion is most suitable for evaluating biases in such complex healthcare systems. This dissertation explores human perception of fairness to identify the most appropriate fairness criterion for assessing AI tools in kidney transplantation, focusing on the preferences of non-expert (e.g. public, patients) stakeholders. The study examines two distinct AI systems employed in kidney transplantation: a classification model and a regression model. Human subject experiments were conducted on the Prolific platform, recruiting 85 participants to rate the fairness of each system independently. These experiments aimed to uncover socially preferred group fairness criteria for evaluating these tools. A Mixed-Logit discrete choice model was employed to analyze fairness feedback, and a projected-gradient descent algorithm is proposed to estimate social fairness preferences. For the classification model, six well-established group fairness notions from the literature were evaluated, with results indicating that accuracy equality is the socially preferred criterion. In contrast, for the regression model—where fairness research is less explored—three novel divergence-based fairness notions were proposed: independence, separation, and sufficiency. Participant preferences revealed a strong inclination toward separation and sufficiency. Additionally, this dissertation introduces five novel mathematical definitions to quantify human perceptions of fairness, leveraging disagreement feedback regarding AI system decisions. By comparing these human-centric measures with actual algorithmic fairness metrics, a notable gap between perceived and computed fairness was observed. This finding prompted an in-depth discussion on strategies to reduce this discrepancy, emphasizing the need to align AI system design with societal fairness expectations.