A Clinically Interpretable Machine Learning Framework for Knowledge Discovery and Early Outcome Prediction in Pediatric Obesity Intervention

Pediatric obesity remains a major public health challenge, yet substantial variability in treatment response limits the effectiveness of standardized intervention programs. Early assessment of short-term weight-change trajectories may provide useful information during an intervention. This study presents a clinically interpretable machine learning framework for predicting short-term pediatric obesity intervention outcomes using routinely collected demographic, anthropometric, body-composition, and early-response measurements from the ProActive Kids program. A systematic workflow including data understanding, clinical data preprocessing, feature engineering, and predictive modeling is developed. Three complementary machine learning paradigms, Logistic Regression, Random Forest, and Extreme Gradient Boosting (XGBoost), are comparatively evaluated to investigate predictive performance and feature importance. Baseline prediction models demonstrate limited predictive performance using only pre-intervention measurements, whereas incorporation of Week-3 early-response information substantially improves prediction across all three classifiers. Week-3 percentage weight change emerges as the dominant predictor of the Week-8 weight-reduction outcome. Waist-to-Hip Ratio, Body Fat Percentage, Body Mass Index, and Fat-Free Mass are also identified as influential baseline variables, although their individual importance should be interpreted cautiously because of correlations among anthropometric and body-composition measures. These findings demonstrate that machine learning can support short-term outcome prediction while identifying potentially clinically relevant patterns in routinely collected pediatric obesity data. The results highlight the predictive importance of early intervention response and may inform future studies of pediatric obesity intervention outcomes.