Metaheuristic-optimized TabNet ensemble for accurate and interpretable obesity classification

Abstract

Obesity is a complex global health issue with severe implications for both individual well-being and public health systems. It has been traditionally challenging to predict and diagnose due to its multifactorial nature, involving genetic, behavioral, and environmental factors. While classical regression models have been extensively used for obesity prediction, their limitations have prompted the exploration of more advanced methodologies. In this study, we leverage Deep Learning (DL) techniques, particularly TabNet, to address the challenges of obesity classification in tabular data—a domain where DL’s potential has often been underutilized. Our approach enhances the TabNet architecture through effective hyperparameter tuning, utilizing Particle Swarm Optimization (PSO), Grey Wolf Optimizer (GWO), and Hunger Games Search (HGS). The resulting models, TabNet-PSO, TabNet-GWO, and TabNet-HGS are combined into a novel ensemble that demonstrates superior performance in obesity classification compared to conventional machine-learning models and recent studies. Additionally, Explainable Artificial Intelligence techniques are employed to provide both local and global interpretability of model predictions, using SHapley Additive exPlanations (SHAP). This interpretability is crucial in clinical settings, where understanding the underlying factors influencing predictions is essential. The study’s findings offer significant contributions to the early detection and management of obesity, providing healthcare professionals with precise and interpretable predictions to guide intervention strategies.