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

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Public Health

Background:

• Obesity is a global health crisis linked to chronic diseases, but traditional classification methods like Body Mass Index (BMI) are insufficient.
• The multifactorial nature of obesity, involving genetics, lifestyle, and environment, complicates accurate categorization and treatment.
• Existing methods struggle to capture complex biological and lifestyle variables, leading to misclassification and ineffective treatments.

Purpose of the Study:

• To develop advanced Artificial Intelligence (AI) models for precise obesity classification.
• To overcome the limitations of traditional BMI-based methods in capturing obesity's complexity.
• To enhance patient stratification for more effective obesity management.

Main Methods:

• Proposed two hybrid AI models: CNN + TabNet and Self-Attention BiLSTM (SA-BiLSTM) + TabNet.
• Employed deep characterization learning combined with TabNet's sparse attention mechanism for feature selection.
• Utilized feature engineering, class-balancing, and Explainable AI (XAI) tools (SHAP, LIME, IG) for model interpretability and robustness.

Main Results:

• CNN + TabNet achieved 92% accuracy (0.925 recall, 0.917 F1-score).
• SA-BiLSTM + TabNet achieved 94.1% accuracy (0.93 recall, 0.937 F1-score).
• Both models demonstrated superior performance compared to conventional baselines.

Conclusions:

• The hybrid models offer a novel approach to joint spatial-temporal learning for interpretable and data-efficient obesity prediction.
• Blending deep characterizations with TabNet's feature selection enhances obesity classification accuracy and clinical utility.
• This AI-driven approach represents a significant advancement over previous hybrid methods and standalone TabNet applications.