Explainable diabetes classification using hybrid Bayesian-optimized TabNet architecture.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Machine Learning for Disease Prediction

Background:

• Diabetes mellitus is a chronic disease with severe health implications if undetected.
• Accurate and explainable early-stage detection is vital for effective diabetes management and treatment.
• Existing diagnostic methods may lack the precision and interpretability needed for widespread adoption.

Purpose of the Study:

• To develop an interpretable machine learning model for accurate early-stage diabetes detection.
• To leverage explainable AI (XAI) techniques for understanding model predictions and identifying key risk factors.
• To enhance user trust and confidence in AI-driven diabetes diagnostic tools.

Main Methods:

• Development of an interpretable TabNet model, optimized using Bayesian optimization (BO).
• Utilization of TabNet's attention mechanism for local and global model interpretability.
• Application of LIME and SHAP XAI tools for robust, model-agnostic explanations.
• Validation using the Pima Indians diabetes dataset (PIDD) and the early-stage diabetes risk prediction dataset (ESDRPD).

Main Results:

• The proposed TabNet model achieved high accuracy: 92.2% on PIDD and 99.4% on ESDRPD.
• Interpretability analysis identified Insulin as the most influential attribute for PIDD (0.301 feature importance).
• Polyuria was identified as the most influential attribute for ESDRPD (0.206 feature importance).

Conclusions:

• The interpretable TabNet model demonstrates superior performance in early diabetes detection compared to benchmark models.
• Explainable AI tools provide crucial insights into the factors driving diabetes classification, enhancing transparency.
• The combination of high accuracy and interpretability is expected to foster greater end-user trust in AI for diabetes screening.