Domain-informed weight forecasting: leveraging behavioral and physiological sequences from wearables

Short-term body-weight forecasting may support personalized weight monitoring, but many existing approaches rely on contemporaneous body weight or body mass index as model inputs, which limits practical use when frequent weigh-ins are unavailable.

We developed a direct multi-step forecasting framework to predict 7-day body-weight trajectories from 14 days of behavioral, physiological, and lifestyle variables. The primary dataset was FitLife360, a synthetic longitudinal dataset on Kaggle, used for model development, benchmarking, and ablation. The LSTM forecaster was evaluated under a participant-level split, with subjects assigned to training, validation, or test sets before window construction. For real-world validation, we tested the framework on PMData from 16 participants over 5 months. Current body weight and BMI were excluded from model inputs and only used as targets. The LSTM was compared with Random Forest, and XGBoost.

On the synthetic FitLife360 dataset, the proposed LSTM achieved the best overall performance in the main comparison and showed consistent gains in the feature-ablation analysis. In the supplementary PMData experiment, the same framework remained operational on real-world wearable/lifelogging records, supporting the feasibility of the approach beyond the synthetic development setting. Detailed metrics for the supplementary experiment are reported in the main text.

These findings should be interpreted primarily as evidence that domain-informed sequence modeling is feasible for short-horizon body-weight forecasting under both controlled synthetic and supplementary real-world data settings. However, such short-horizon predictions should not be interpreted as direct measures of meaningful adiposity change or short-term cardiometabolic risk, because day-to-day body weight also reflects transient physiological variability. The study therefore provides a methodological foundation for future validation in larger real-world and clinical cohorts.