Abstract Introduction Accurate weight fluctuation monitoring is critical for managing cardiometabolic health, yet reliance on manual scales undermines adherence and scalability. Smart bed technology eliminates this barrier by passively tracking weight from overnight pressure signals, requiring minimal user effort. This study develops an Extreme Gradient Boosting (XGBoost)-based framework to detect clinically relevant weight changes (≥2.5%) and longitudinal trends, enabling seamless integration of weight monitoring into daily life while maintaining the accuracy necessary for meaningful health surveillance. Methods Six participants collected smart bed pressure data over 2 weeks, with daily reference weights from a weighing scale.Smart bed pressure signals were downsampled from 1 kHz to 1 Hz to reduce noise. Data were segmented into in-bed and baseline periods, and signal stability was assessed using entropy-based metrics on 30-second windows. Stable windows (low entropy) were retained, and median pressure values were extracted for analysis. Feature engineering utilized two data sources: raw pressure signals and high-quality in-bed readings from the first 60 seconds post-entry. Nightly summaries (mean and median) were standardized by subtracting baseline pressure. Relative changes were expressed as percentages. The two most predictive features were standardized nightly means computed with and without quality filtering. Temporal trends were derived by calculating weekly weight-change slopes over seven-night smoothened windows. Features were rescaled using min–max normalization before XGBoost training. The optimal model used a generalized linear booster with 40 estimators, a maximum depth of 5, learning rate of 0.5, and ridge regularization of 0.01. Subject-level leave-one-out cross-validation evaluated model performance using coefficient of determination (R²), Bland–Altman limits of agreement (LoA) for weight change, and confusion matrix metrics for trend classification at a 2.5% threshold. Results Weight-change estimation showed strong agreement with reference measurements (R² = 0.91; LoA ±0.12 lbs./night; mean bias 0 95% CI: ±0.02). Trend classification achieved 83% accuracy with sensitivity 0.82, specificity 0.85, precision 0.88, and Cohen’s kappa 0.66, demonstrating robust detection of weight trajectories. Conclusion This approach enables seamless at-home monitoring of weight fluctuations and longitudinal trends, supporting scalable personalized health tracking with minimal user burden. Limitations include small sample size and associated epistemic uncertainty. Larger studies are needed to confirm generalizability. Support (if any)
Rao et al. (Fri,) studied this question.