International Journal of Science, Engineering and TechnologyAuthors: Er. Madeeha Laiq, Er. Animesh Kushwaha, Er. Ajaz Ahmad
Abstract: Early identification of emerging weight gain may enable preventive intervention before a sustained increase becomes established. Conventional weight-gain prediction often uses static characteristics or long prediction horizons, while short-horizon forecasting primarily estimates future numerical weight. This study proposes AWTG-PD (Adaptive Weight Trajectory Guidance with Personal Deviation), a longitudinal machine-learning framework combining multi-window weight trajectories with deviation from an individual-specific recent baseline. The study used the publicly available FitLife360 health and fitness tracking dataset obtained from Kaggle. Weight features included 7-days, 14-days, and 30-day averages and changes, trajectory slope, personal-baseline deviation, and deviation z-score. A near-term gain event was defined as a future 14-day average weight increase of at least 0.10 kg. Chronological training, validation, and test periods were used, followed by ablation, threshold, early-warning, risk-stratification, bootstrap, and calibration analyses. On the held-out test set, AWTG-PD achieved ROC-AUC 0.7018 (95% CI: 0.6745–0.7263) and PR-AUC 0.1818 (95% CI: 0.1537–0.2144). At threshold 0.65, precision was 0.1914, recall 0.3281, F1 0.2418, and alert rate 16.44%. The model generated early warnings for 124 users with a mean lead time of 9.008 days. At the same alert rate, a 30-day baseline generated 119 early-warning users with a mean lead time of 8.807 days. Risk stratification identified a high-trajectory/high-deviation group with a gain-event rate of 19.76% and relative risk of 2.459. A lifestyle-only model achieved ROC-AUC 0.5050. Calibration indicated substantial overestimation of absolute risk, so model outputs should currently be interpreted as relative alert scores rather than calibrated probabilities. The findings support trajectory-based early warning as a methodological approach, while external validation in real-world longitudinal cohorts remains necessary.
