Explainable Hybrid Deep Learning–Transformer Approach for Insulin Prediction

Accurate predictive modeling is critical for enhancing patient outcomes and facilitating personalized care. This study introduces a hybrid modelling framework that combines deep learning, transformer-based architectures, and classical regression methods. The framework integrates multiple approaches, including Artificial Neural Networks, Long Short-Term Memory Networks, Convolutional Neural Networks, Random Forest, to model complex patterns in insulin biomarker data. By integrating these models into a unified framework, the approach enhances predictive accuracy while ensuring interpretability. Explainable AI techniques, including SHAP and LIME, are employed to identify key features influencing predictions, thereby promoting transparency and clinical trust. The proposed framework achieves superior performance on clinical datasets, with improved metrics such as MSE, MAE, and R², outperforming baseline models. Additionally, it identifies critical biomarkers associated with insulin regulation. Subgroup-level interpretations provide clinically relevant insights that inform personalized treatment strategies. This work demonstrates how advanced machine learning, coupled with explainability, establishes a robust foundation for clinical decision support systems to deliver effective and individualized patient care.