Longitudinal prediction of BMI using explainable AI: integrating polygenic scores, maternal, early-life and familial factors

Abstract Background/objectives This study aimed to predict body mass index (BMI) trajectories from childhood to early adulthood using explainable artificial intelligence, integrating adult BMI polygenic scores (PGS), maternal, early-life, and familial factors to identify key predictors of obesity risk and inform prevention strategies. Subjects/methods We analyzed longitudinal data from the Raine Study Gen2 cohort, recruiting 2868 participants. This observational study, without randomization or case-control design, collected BMI measurements at ages 8, 10, 14, 17, 20, 23, and 27 years. We applied Kolmogorov–Arnold Networks (KAN) alongside conventional machine learning models, integrating epidemiological variables (maternal and paternal anthropometrics, parental education, early-life skinfold measurements) with seven BMI-related PGS. The analysis spanned from childhood to early adulthood, with no intervention administered. Results The KAN model, combining epidemiological and PGS data, achieved predictive performance with R² ranging from 0.81 for BMI at age 8 to 0.34 at age 27. BMI z-score at age 5 was the dominant predictor in early years, with adult BMI PGS influence increasing post-adolescence. Maternal and paternal anthropometry, parental education, and early-life skinfold measurements were significant contributors. Conclusions The interpretable KAN model revealed the dynamic interplay of childhood BMI z-score and PGS emerging as key drivers of BMI trajectories across life stages. The finding underscores the potential of BMI at critical time in early childhood as a biomarker for obesity risk. Our interpretable model offers actionable insights for targeted obesity prevention strategies.