Cost Sensitive Interpretable Decision Making Framework for Diabetes Prediction

Abstract

Machine learning in diabetes screening faces two challenges: data imbalance, leading to missed diagnosis of high-risk patients, and the lack of clinical interpretability of the "black box" model. This paper proposes an interpretable Light Gradient Boosting Machine (LightGBM) prediction architecture based on cost sensitive and Bayesian joint optimization. In this study, the core parameters of the model and the penalty weight of the positive sample cost are synchronously introduced into the Bayesian automated optimization (Optuna) search space for integrated optimization. The empirical study based on Behavioral Risk Factor Surveillance System (BRFSS 2015) data set (253680 samples, positive samples accounted for 13.93%) shows that this model has achieved a Area Under the Receiver Operating Characteristic Curve (ROC-AUC) score of 0.8259 and a recall rate of 0.8281, effectively controlling the missed diagnosis rate. In addition, the architecture integrates Tree SHapley Additive exPlanations (TreeSHAP) algorithm, converts the nonlinear decision boundary into visual indicators, and identifies the three core signs of disease: general health, hypertension and age. Under the premise of preserving the characteristics of the original data, the effective combination of prediction accuracy, missed diagnosis control and clinical interpretability is realized, which provides an auxiliary decision-making scheme for chronic disease intervention.