An Early Diabetes Detection Framework Utilizing Interpretable Hybrid Deep Learning Model

Diagnosing diabetes, a chronic metabolic disorder distinguished by heightened levels of glucose in the blood, presents substantial obstacles and presents considerable health hazards. Detection of diabetes that is both accurate and efficient is vital for timely intervention and management. Utilizing a binary tabular dataset, we present an innovative interpretable hybrid deep learning model proposed in this study. Our methodology incorporates a stacked ensemble model (EM) that consists of Gradient Boosting Machine (GBM), Random Forest (RF), and Naive Bayes (NB) as base-classifiers, in addition to an Artificial Neural Network (ANN) functioning as the meta-classifier. By capitalizing on the respective merits of conventional machine learning (ML) algorithms and deep learning methodologies, this amalgamation seeks to improve the precision of predictions and the resilience of models. In order to effectively address the crucial requirement for interpretability in medical diagnostic models, we have utilized SHAP (SHapley Additive exPlanations) values. These values provide valuable insights into the decision-making process of our model and highlight the specific contribution of each feature to the final prediction outcome. The effectiveness of the proposed model in detecting diabetes has been substantiated through a comprehensive evaluation of its performance utilizing a range of metrics. The proposed model has achieved 99.22% accuracy.