Paul Festor

Paul Festor

3 previous affiliations
    • PhD student · 2020–2024
    • Graduate Teaching Assistant · 2023–2024
    • Graduate Teaching Assistant · 2021–2024
    • Course Instructor · 2022
    • MSc in Computing (Artificial Intelligence & Machine Learning), Intelligence artificielle · 2019–2020
  • Vocads2021–2023
    • AI Science Communicator · 2021–2023
  • efluid SAS2019
    • Data Science Intern · 2019
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Paul Festor focuses on explainable artificial intelligence and reinforcement learning within clinical decision support systems. His work investigates how medical professionals interact with algorithmic recommendations, utilizing high-fidelity simulations and eye-tracking technology to evaluate clinician behavior. Recently, he has examined the safety implications of these systems, particularly how doctors respond to safe versus unsafe algorithmic suggestions in critical care scenarios.

Research Trajectory

The deployment of artificial intelligence in healthcare has historically been slowed by safety concerns and a lack of empirical understanding regarding how expert clinicians interact with algorithmic advice. Much of the early evaluation of explainable AI (XAI) relied on non-expert users in simplified, low-fidelity settings. Festor’s research addresses this gap by bringing XAI evaluation into high-fidelity clinical environments, focusing specifically on the behavioral dynamics between intensive care doctors and decision support tools.

In 2023, Festor collaborated on foundational studies utilizing eye-tracking technology to measure clinician attention during simulated medical tasks. As a co-author of Eye-tracking of clinician behaviour with explainable AI decision support: a high-fidelity simulation study, he helped design and analyze a prescription dosing scenario in a simulation suite. This work sought to determine whether XAI could serve as an effective safety mitigation mechanism, particularly in the era of generative AI where hallucinatory or unsafe suggestions pose severe risks to patient harm. By tracking where and how long doctors looked at the provided explanations, the research provided quantitative behavioral readouts of human-AI interaction.

This methodological approach was concurrently applied to reinforcement learning (RL) systems, a critical technology for sequential medical interventions. In Interaction of doctors with explainable RL decision support via behavioural readouts of eye-tracking, Festor and his colleagues evaluated explainable RL (XRL) in the same high-fidelity dosing scenarios. This work intersects with the research direction of Learning Optimal Sequences of Medical Interventions, emphasizing that the utility of an RL policy in a clinical setting depends heavily on whether the human operator can safely interpret its outputs.

In 2024, Festor took a lead role in advancing this specific analytical framework. As the first author of Evaluating the impact of explainable RL on physician decision-making in high-fidelity simulations: insights from eye-tracking metrics, he further detailed how cognitive attention is allocated when clinicians are presented with safe versus unsafe XRL suggestions. The findings from this line of inquiry demonstrate that the cognitive attention devoted to XRL during unsafe scenarios is measurable and distinct, providing a pathway to better understand how doctors verify or reject algorithmic advice. This empirical approach to AI safety aligns with the broader area of Structured Safety Cases for High-Risk Systems, grounding theoretical safety concerns in observable human behavior.

Current Directions

The most recent papers in the evidence address the practical and safety-critical challenges of deploying explainable reinforcement learning in high-stakes medical environments. Recent work investigates how expert clinicians allocate cognitive attention when interacting with AI decision support, particularly when algorithms provide unsafe or hallucinatory recommendations. By bridging clinical simulation with behavioral metrics like eye-tracking, this research aims to empirically evaluate whether explainability mechanisms can effectively mitigate safety risks and improve human-AI collaboration in critical care.

Profile updated

See something wrong? Report a mistake