Interpreting Type 1 Diabetes Management via Contrastive Explanations

Advances in Type 1 Diabetes (T1D) management technologies have led to major life improvements for children and adults affected by this condition. Among other approaches, Reinforcement Learning (RL) algorithms play an essential role in the design process. Indeed, applications of T1D management such as Actor-Critic methods or Q-Learning have shown great success in this regard. Sometimes, however, the precise explanations behind why an agent chooses different actions requires major cognitive efforts by the end users. To address this, explainability techniques such as Decomposed Q-values (DQv), Reward Difference Explanations (RDX), and Belief Maps (BM) built via intended outcomes can help better understand the process. As such, results obtained via questionnaires and votes show that explainable RL (XRL) procedures help clarify the steps that the controller evaluates before acting. Tests were done on a benchmark dataset and then presented to 19 patients and 5 physicians. On average, users have expressed the preference towards visual explanations, such as BMs, instead of the other types of explanations analyzed. This suggests that the use of explainability for T1D management is a crucial step in obtaining better insights and increased trust towards AI systems for both patients and experts in the field.