Longitudinal Dietary Optimization via Multi-Horizon Time Series Forecasting of BMI with Transformer Networks and Personalized Recommendation using Collaborative Filtering with Implicit Feedback

Abstract

Long-term management of Body Mass Index (BMI) shows to be challenging for most people, mainly due to limitations in dietary compliance monitoring and the lack of personalized predictive guidance. Standard methods leave much to be desired when it comes to offering personalized, adaptable strategies aligned with our unique metabolic processes and changing preferences. To address this gap, this paper develops a new computational framework for longitudinal dietary optimization. We leverage early detection for BMI management as our system is powered by state-of-the-art machine learning algorithms that provide tailored, time-sensitive dietary recommendations for proactive BMI control.

The proposed system is composed of two key components that work interdependently with each other. Firstly, we proposed a multi horizon time series forecasting model of transformer networks. This element is crucial in this context as it allows us to model how a user timeline, given all their past nutritional data, can still behave differently given food ordering tendencies. By learning from these multi-axis sequences, the Transformer model predicts BMI at different future time points, allowing for preemptive rather than reactive dietary changes. Next we use CF) with implicit feedback to deploy an personalized food recommendation engine. This CF module treats users' past food order histories as implicit indicators of preference as a user. It then looks for patterns and similarities from user to user to identify foods which not only align with an inferred taste profile for that individual, but are also preferred by other, similar users.

A set of key inputs underpin the operation of the system, including longitudinal user data (history of BMI and food ordered logs) along with a database of foods with detailed nutritional information for each food item (e.g., calories, macronutrients, micronutrients sourced from datasets such as USDA FoodData Central). The main outputs are a multi-horizon BMI forecast and a dynamically generated optimized meal plan. This is designed based on anticipated BMI trajectory, the user-specific preferences from the CF module, and preset nutritional goals; For example, when BMI is predicted to increase, the plan would suggest lower-calorie, high-fiber foods, whereas for maintenance it would promote protein-rich foods. The system can be augmented with cautionary information, for example on excessive calories or fat content.

This work primarily focuses on a novel approach by optimizing state-of-the-art multi-horizon time series forecasting to directly personalize recommendations for longitudinal dietary optimization. Our system has enormous potential to support users in understanding how their BMI deviations from the normal range may influence future weight and to guide them with personalized food recommendations, leading to healthier patterns and long-term health drivers. This technology is an advancement in personalized nutrition driven by data..