TOWARDS IMPROVING DIABETES CLASSIFICATION USING DEEP LEARNING BY BALANCING ACCURACY AND INTERPRETABILITY

Abstract

Diabetes is a chronic condition that necessitates early prediction and protection to prevent serious complications. Although existing machine learning models for diabetes classification often prioritize predictive accuracy, they frequently lack scalability and interpretability. This lack of transparency and adaptability limits their applicability across heterogeneous patient populations. To address these challenges, this study proposes a novel Interpretable Multi-head Attention Deep Learning with SHAP-based Interpretability (IMHA-DLSI), which integrates multiple complementary techniques. Specifically, this framework incorporates IA- TabNet-FS inspired with features selection SHAP (SHapley Additive exPlanations) and an Interpretable Multi-head Attention Artificial Neural Network (IMHA-ANN) as the core predictive model. The core model was trained on 80% and tested on 20% of datasets (PIMA-IDD, DDFH-G) within the IMHA-DLSI framework to ensure robust and comprehensive performance, achieved exceptional results across these datasets. The proposed model attained 80.52% accuracy on training dataset and 81.19% accuracy on unseen data of PIMA-IDD. Similarly this model attained 98.94% on training dataset and 98.51% on unseen dataset of DDFH-G. Model predictions were rigorously traced with SHAP, demonstrating that the proposed IMHA-DLSI framework significantly outperforms existing models in diabetes detection. Furthermore, it incorporates dynamic threshold optimization and exhibits strong Gaussian noise (σ = 0.0005) for prediction stability, enhancing its readiness for clinical deployment. This work addresses critical limitations in current ML-based diabetes classification methods by offering a transparent, scalable, and high- performing solution. The IMHA-DLSI framework thus represents a significant advancement in the application of interpretable artificial intelligence for precision healthcare.