Interpretable Machine Learning Framework for Diabetes Prediction: Integrating SMOTE Balancing with SHAP...

Area of Science:

• Machine Learning in Healthcare
• Diabetes Prediction Models
• AI for Clinical Decision Support

Background:

• Class imbalance and limited interpretability hinder clinical adoption of AI for diabetes prediction.
• Existing models often show poor sensitivity to high-risk cases and lack clinical trust.
• This study integrates SMOTE resampling with SHAP explainability to improve performance and transparency.

Purpose of the Study:

• To develop and validate an interpretable machine learning framework for diabetes prediction.
• To address class imbalance using advanced resampling techniques (SMOTE).
• To provide clinically meaningful explanations via SHAP for enhanced decision support.

Main Methods:

• A seven-stage pipeline combining Synthetic Minority Oversampling Technique (SMOTE) and SHapley Additive exPlanations (SHAP).
• Comparative evaluation of five algorithms (Random Forest, Gradient Boosting, SVM, Logistic Regression, XGBoost) on 1500 patient records.
• 5-fold stratified cross-validation with SMOTE applied within training folds to prevent data leakage.

Main Results:

• The Random Forest-SMOTE model achieved 96.91% accuracy, 0.998 AUC, 99.5% sensitivity, and 97.3% specificity.
• SHAP analysis identified glucose (SHAP value: 2.34) and BMI (SHAP value: 1.87) as primary predictors.
• Feature interaction analysis revealed synergistic effects between glucose and BMI.

Conclusions:

• The proposed framework shows promise in bridging algorithmic performance and clinical applicability for diabetes prediction.
• The Random Forest-SMOTE model demonstrated high cross-validated performance on a publicly available dataset.
• Further external, prospective validation in real-world cohorts is required before clinical deployment.