A Differential Privacy and TOPSIS Enhanced Explainable Machine Learning Framework for Diabetes Risk Diagnosis

Diabetes is a chronic condition affecting blood sugar regulation and impacts a significant portion of the global population. Early detection is crucial, as it can help reduce complications and improve health outcomes. Many cybercrime attacks are directed toward the healthcare sector, underscoring the importance of secure data handling. In this study, we use a dataset to predict diabetes risk, employing Machine Learning (ML) which offers a powerful means for accurate prediction by leveraging complex patterns in health data, yet privacy concerns around sensitive medical information remain a significant challenge. This study addresses this concern by incorporating Differential Privacy (DP), specifically utilizing the Laplacian Mechanism (LM), to protect patient data. We employ a range of ML algorithms, including Extreme Gradient Boosting (XGB), Random Forest (RF), Gradient Boosting Decision Trees (GBDT), Bootstrap Aggregating (Bagging), and Stacked Generalization (Stacking), to ensure robust model performance. Using the Technique for Order Preference by Similarity to the Ideal Solution (TOPSIS) for statistical analysis, our results reveal that even under DP constraints, the XGB model achieves an impressive accuracy of 89.43% while providing superior privacy protections. In contrast, without DP constraints, the RF model reaches a higher accuracy of 98.27%. To enhance interpretability, we integrate Explainable Artificial Intelligence (XAI) techniques such as Shapley Additive explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), allowing us to understand the influence of individual features on the model's predictions. Our study also employs 10-fold cross-validation to confirm the model’s stability and reliability. This approach not only supports accurate and private diabetes prediction but also paves the way for the application of DP in broader healthcare ML applications, balancing data privacy with predictive utility.