Effective Machine Learning model for Diabetes insulin sensitivity

Abstract

Accurate assessment of insulin sensitivity is essential for personalized diabetes management; however, conventional clinical methods are often invasive, time-consuming, and impractical for large-scale screening. This study proposes a machine learning–based framework for estimating personalized Insulin Sensitivity Index (ISI) values from routinely collected clinical variables. A two-layer stacking ensemble architecture was developed by combining three heterogeneous base learners: Artificial Neural Network (ANN), XGBoost, and LightGBM, with a Ridge Regression meta-learner. The framework was trained and evaluated using the Pima Indians Diabetes Dataset containing 768 patient records. Since direct ISI measurements were unavailable, a validated QUICKI-inspired proxy formula based on fasting glucose and fasting insulin was used to derive the target regression variable. Missing and physiologically implausible zero values were handled using median imputation, and model performance was assessed through 5-fold cross-validation. The proposed stacking ensemble achieved superior predictive performance with R² = 0.9845, RMSE = 3.88, and MAE = 1.46, outperforming linear regression, ANN, LightGBM, and XGBoost when used individually. Feature importance analysis identified insulin and glucose as the most influential predictors, while age and body mass index provided additional discriminatory information. Furthermore, the estimated ISI values demonstrated strong biological consistency, with lower ISI observed in diabetic individuals and progressive decline across increasing age and BMI groups. These findings indicate that the proposed framework offers an accurate, interpretable, and non-invasive approach for personalized ISI estimation and provides a practical foundation for future ODE-based glucose–insulin modeling and AI-driven digital twin systems for diabetes management.