Reinforcement Learning-Driven Nutrition Coaching

This paper presents an AI-enabled nutrition coach that combines computer vision with reinforcement learning (RL) to support diabetes self-management. Deployed as web and mobile applications, Dietra analyzes meal photos to estimate calories and macronutrients and adapts coaching based on user behaviour and feedback. In an initial deployment with 20 adults with Type 1 or Type 2 diabetes, the system achieved 90.4% accuracy for calories, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 2 \%}$</tex> for carbohydrates, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 3 \%}$</tex> for protein, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 9 \%}$</tex> for fat against internal ground truth. An RL agent updates meal plans and nudges to optimize day-, week-, and month-level adherence. A dashboard displays calorie/macronutrient totals, a nutrition score, streaks, and real-time recommendations. Early results suggest that the adaptive feedback loop improves dietary awareness and supports adherence, motivating a larger clinical evaluation.