Why understanding AI matters: explainable reinforcement learning in healthcare for type 1 diabetes management
Author(s)
Melloni, Daniele
Date Issued
November 27, 2025
Type
Doctoral Thesis
Abstract
Artificial Intelligence (AI) is becoming increasingly central in healthcare, yet its widespread
adoption remains hindered by the lack of transparency of complex models often perceived
as “black boxes.” In this context, explainability (XAI) emerges as a key requirement to
ensure trust, safety, and understanding of automated decision-making processes.
This thesis investigates the role of explainability in Reinforcement Learning (RL) systems
applied to Type 1 Diabetes Mellitus (T1DM) management. Following an extensive review of
the current literature, the work explores the use of RL and Deep RL techniques in this
domain, with a specific focus on interpretability challenges.
Three main applications of Explainable Reinforcement Learning (XRL) are presented: (1)
construction of counterfactual explanations from an actor–critic framework; (2) derivation
of post-hoc explanations through Structural Causal Models built from a Deep Q-Network;
and (3) generation of visual explanations using Markov networks extracted from the same
DQN.
The results show that these methods substantially improve the interpretability and clarity
of the models, making them more transparent to both clinical experts and laypeople. In
conclusion, this work highlights how integrating explainability techniques within RL
frameworks represents a crucial step toward clinically reliable, interpretable, and humancentered
AI.
Additional information
Dottorato di ricerca in Engineering for Energy and Environment
File(s)![Thumbnail Image]()
Name
dmelloni_tesid.pdf
Size
2.65 MB
Format
Adobe PDF
Checksum (MD5)
d419e41b5b8e644755174dfffe0fea47