Maintaining normal glucose levels through dietary and lifestyle behaviors can help with losing weight, optimizing mental health, suppressing food cravings, improving sleep, managing diabetes, and reducing risks of developing cardiovascular disease. Dysglycemia can lead to chronic complications such as diabetes, kidney disease, myocardial infarction, stroke, amputation and even death. Tools to predict and offer immediate actionable feedback, instead of broad lifestyle recommendations, can help mitigate many of these complications.
Explainable AI can provide insights into a model’s decision-making process and offer guidance. One such example is counterfactual explanations, which provide insights into why a model makes a particular prediction by generating hypothetic instances that are similar to the original input but lead to a different prediction. However, integrating user preferences into the explanations remains an open research problem to designing AI-driven health interventions.
Researchers at Arizona State University have developed a novel framework, GlyCoach, for generating counterfactual explanations to control blood glucose. GlyCoach characterizes the decision boundary for high-dimensional health data and is able to generate actionable interventions from a grid search. A unique feature of GlyCoach is that it integrates prior knowledge about user preferences of plausible explanations into the process of counterfactual generation. When evaluated using real-world datasets and external simulators, GlyCoach achieved 87% sensitivity, surpassing state of the art techniques for generating counterfactual explanations. Further, the counterfactuals from GlyCoach demonstrate a 32% improvement in normalized distance.
Potential Applications
- Blood glucose control (could be part of a mobile app with integrated machine learning/AI algorithms and wearable sensors to prevent dysglycemia)
- Provides actionable feedback for behavioral changes
Benefits and Advantages
- Provides immediate actionable recommendations to the user to prevent events that occur in the short-term (such as hyperglycemia after eating)
- Uses a grid search approach to find counterfactual explanations that align with users’ preferences
- Users may make behavioral changes based on preference (e.g. reducing carb vs meat consumption)
- An auto-encoder model is used to generate adversarial examples which are then used to construct the decision boundary
- More tailored toward preventing glucose excursions
For more information about this opportunity, please see
For more information about the inventor(s) and their research, please see