Abstract Childhood obesity is a complex and multi-factorial condition influenced by genetic predisposition, environmental exposures, and early-life anthropometrics. While machine learning (ML) models have shown promise in predicting obesity trajectories, their adoption in clinical settings is limited due to a lack of interpretability. In this study, we apply Kolmogorov-Arnold Networks (KAN), an explainable deep learning framework, to predict body mass index (BMI) as an obesity risk indicator at 8 years old, using early-life epidemiological factors and polygenic risk scores (PGS) from the Raine Study Gen2 cohort. KAN’s formularization mechanism enables mathematical representation of predictive relationships, allowing for improved model transparency. Our results demonstrate that KAN outperforms traditional ML models—including Extreme Random Forest, XGBoost, and Lasso regression—achieving the highest R2 score (0.81) when integrating polygenic scores and epidemiological data. Feature importance analysis identifies Body Mass Index (BMI) z-score at Year 5, triceps and suprailiac skinfold thickness, and polygenic scores as key contributors. These findings highlight the potential of explainable deep learning in personalized obesity prevention, providing a transparent and interpretable AI-driven tool for early intervention strategies.
Chen et al. (2025) studied this question.