An Integrated Predictive, Explainable, and Causal Machine Learning Framework for Early Detection of Type 2 Diabetes with Uncertainty and Fairness Assessment

Abstract

Early detection of Type 2 Diabetes Mellitus (T2DM remains a critical challenge in preventive healthcare due to the complex interplay of clinical, demographic, and lifestyle factors. This study proposes an Integrated Predictive, Explainable, and Causal Machine Learning Framework for early detection of Type 2 diabetes, incorporating uncertainty quantification and fairness analysis to enhance clinical reliability and generalizability. The proposed framework is evaluated using the Pima Indians Diabetes Dataset, comprising clinical measurements such as glucose concentration, body mass index, blood pressure, and age. An ensemble predictive architecture centered on XGBoost achieves superior classification performance, attaining an accuracy of 87.4% and a ROC-AUC of 0.91 under stratified 10-fold cross-validation. Model transparency is ensured through SHAP-based explainability, enabling clinicians to identify dominant risk contributors, particularly plasma glucose levels and BMI. To move beyond correlation-based inference, a Structural Causal Model is employed to estimate the causal influence of modifiable risk factors on diabetes onset. Additionally, fairness-aware evaluation is conducted using demographic parity and equal opportunity metrics, revealing reduced prediction bias across age and gender subgroups after fairness constraints are applied.