An XGBoost-based predictive framework for diabetes mellitus multi-classification

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Biology

Background:

• Diabetes mellitus is a chronic metabolic disorder requiring early detection to mitigate severe health consequences.
• Accurate multi-class classification (non-diabetic, pre-diabetic, diabetic) is crucial for timely intervention.
• Existing predictive models may lack the precision needed for complex diabetes stratification.

Purpose of the Study:

• To develop and evaluate an Extreme Gradient Boosting (XGBoost)-based predictive framework for multi-class diabetes mellitus prediction.
• To compare the performance of XGBoost against other machine learning algorithms for diabetes classification.
• To assess the impact of different data preprocessing and resampling techniques on predictive accuracy.

Main Methods:

• A dataset of 826 unique clinical records was preprocessed, including standardization and duplicate removal.
• A stratified 70:30 train-test split and five-fold cross-validation were employed.
• Extreme Gradient Boosting (XGBoost) was implemented and benchmarked against Logistic Regression, Random Forest, Support Vector Machine, Decision Tree, and K-Nearest Neighbors.

Main Results:

• The XGBoost model achieved a high accuracy of 99.60% on the original dataset.
• Resampling techniques like Random Over Sampling (ROS) and Synthetic Minority Over-sampling Technique (SMOTE) also yielded 99.60% accuracy, improving generalization but increasing computational cost.
• Under-sampling methods (Random Under Sampling, Cluster Centroids) resulted in lower accuracies (92.74%, 90.32%).
• XGBoost demonstrated superior performance compared to all other benchmarked algorithms, with an average cross-validation accuracy of 98.79% ± 1.13%.

Conclusions:

• The proposed XGBoost framework is highly effective for accurate multi-class diabetes mellitus prediction.
• XGBoost outperforms traditional machine learning algorithms in this classification task.
• The study underscores the importance of robust predictive modeling for early diabetes diagnosis and management.