Adopting Post-Hoc Explainable Reinforcement Learning in Healthcare Scenarios
Daniele Melloni, Andrea Zingoni · 2025
As artificial intelligence systems become predominant in multiple applications, the transparency and trustworthiness of models are not prioritized enough. In this work, a novel application of Markov networks is explored in the healthcare sce-nario of Type 1 Diabetes Mellitus (TIDM) management. Attached to the Deep Q-Network (DQN) agent, the Markov networks are able to increase transparency in the inner mechanics of the system. This results in an improved trust and accountability of the algorithm, enhancing human-AI collaboration and decision support. In particular, explanations generated by this method resulted in a 28% increase in trust towards the closed-loop system when presented to 41 real TIDM patients. In conclusion, using post-hoc explainability methods such as Markov networks in applications such as healthcare is of paramount importance for the adoption of safety-critical agents.