An Interpretability-Driven Fuzzy Modeling Methodology for Personalized Meal Detection

In the self-management of Diabetes, meals are an important factor as they dictate how the blood glucose (BG) levels vary. Efforts on creating systems that aid diabetic people to rely on processes like meal detection. Such systems give priority to be accurate while sacrificing interpretability. Fuzzy systems represent knowledge as close as possible to natural language in an effort to increase interpretability. However, many approaches are still solely data-driven, putting aside the domain knowledge input, and not guaranteeing interpretability in the final model. In this paper, we propose a fuzzy modeling methodology mainly focused on interpretability. The proposed methodology allows for a domain expert to tailor the model structure in order to determine the interpretable boundaries. Within such boundaries, personalized data-driven models are developed considering the interaction with the end user, the integration of domain knowledge, and the trust between end user and system. We apply the methodology to a meal detection scenario, giving examples of how the models' output could be used to create interpretable information.