Health Prediction Using Hybrid Deep Learning Models: A Transparent and Interpretable Approach

Abstract

Obesity has become alarming globally, with mounting health emergencies related to a number of chronic conditions like cardiovascular disease, diabetes, and metabolic disorders. Despite growing global awareness, accurately categorising obesity remains challenging due to its multifactorial nature involving genetic, lifestyle, environmental, and psychological determinants. Traditional methods based on pristine Body Mass Index (BMI) fail to encapsulate the multifaceted biological and lifestyle variables responsible for weight gain, leading to misclassification and treatment unresponsiveness. To overcome the challenge of a precise obesity classification, this study proposes two hybrid Artificial Intelligence (AI) models—CNN + TabNet and Self-Attention BiLSTM (SA-BiLSTM) + TabNet—that conjointly employs deep characterisation learning and TabNet's sparse attention–mechanism feature selection. To ensure better comprehension, the proposed models use feature engineering, class-balancing, and Explainable AI (XAI) tools, which include SHAP, LIME, and Integrated Gradients (IG). The performed experiments exhibit improved and better results over other conventional baselines. Viz. CNN + TabNet is 92% correct (recall 0.925, F1 0.917), while SA-BiLSTM + TabNet is 94.1% correct (recall 0.93, F1 0.937). The novel approach of blending TabNet's dynamic sparse-attention feature selection with CNN- and SA-BiLSTM-based deep characterisations enables the joint spatial-temporal learning with interpretable and data-efficient obesity prediction. Former hybrid approaches either combined generic deep networks (like CNN-LSTM and DNN-RF) or used TabNet on its own.