Personalized Glucose Management With AI: Pilot Study Using a Multiarmed Bandit Approach.

Area of Science:

• Digital Health
• Behavioral Science
• Metabolic Health

Background:

• Mobile apps offer personalized behavioral recommendations for chronic disease prevention, notably diabetes.
• Current reinforcement learning approaches optimize recommendations but often neglect clinical outcomes.
• Focusing solely on behavior change overlooks critical clinical results.

Purpose of the Study:

• To propose an online planning method for dietary and exercise recommendations.
• To optimize postprandial glucose levels directly through behavioral modifications.
• To enhance clinical outcomes by linking behavioral changes to glucose response.

Main Methods:

• A multiarmed bandit approach with a two-stage reward prediction model was developed.
• Actions involved combinations of carbohydrate intake and postprandial walking duration.
• Rewards were defined as reductions in postprandial glucose levels, incorporating predicted behavioral and glycemic responses.

Main Results:

• Simulation experiments demonstrated significant improvements in postprandial glucose levels compared to randomized policies.
• A small-scale real-world experiment with 6 participants showed a 23% average improvement in glucose response.
• High behavioral adherence to carbohydrate intake and walking recommendations was observed.

Conclusions:

• The proposed method shows preliminary effectiveness in improving postprandial glucose control through personalized recommendations.
• Both simulation and real-world experiments support the method's potential.
• Further longitudinal studies in diabetic patients are necessary for validation and generalization.