Explainable stacking ensemble with feature tokenizer transformers for men’s diabetes prediction

Abstract

Diabetes is a leading global health concern, with millions of deaths linked to diabetes and related complications according to the World Health Organization (WHO). Early and accurate prediction is crucial for effective management. This study investigates the potential of a stacking ensemble approach for predicting diabetes in men (n = 5598). The ensemble leverages a Feature Tokenizer transformer, a deep learning technique, alongside various machine learning models. SHAP (SHapley Additive exPlanations) is used to enhance model interpretability. Compared to other stacking methods and standalone models, the proposed ensemble with a Random Forest meta-classifier, XGBoost, Feature Tokenizer Transformers (FT-Transformer) and LightGBM achieved superior performance (accuracy: 0.8786, precision: 0.7989, recall: 0.8171, F1-score: 0.8079, Area Under the Curve (AUC): 0.8618). These findings suggest that stacking ensembles with deep learning and explainable artificial intelligent (AI) hold promise for improving diabetes prediction in men, potentially leading to better clinical decision-making and patient outcomes.

Keywords

• feature tokenizer
• men’s health
• diabetes
• explainable artificial intelligent