A comparative and interpretable machine learning framework for reliable diabetes risk prediction.

Abstract

Diabetes mellitus is a common type of metabolic illness that is very common worldwide, and in most cases, it results in serious effects like heart disease, kidney disease, and blindness. Proper and early diagnosis of diabetes is essential to intervene on time and have better patient outcomes. Machine learning (ML) paradigms provide effective predictive modeling solutions to healthcare, but most of the current literature is limited due to imbalanced datasets, using a single training test split, and limited model interpretability, which diminish their clinical usability. This research paper has introduced a powerful and explainable ML model to predict diabetes based on the Pima Indians Diabetes Dataset acquired via Kaggle, which contains 768 patients with eight clinical variables and a binary response. To counter the class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is used to create natural synthetic samples of the minority diabetic group that facilitate balanced learning without degrading the correlations between the features. Four classifiers, including Logistic Regression, Naive Bayes, AdaBoost, and XG Boost, are trained and tested. The stratified 10-fold cross-validation is used to provide a stable and generalizable model performance, as opposed to using only one data split. The measurement criteria are accuracy, precision, recall, and F1-score, especially for the minority diabetic class. The interpretation of the model is improved by the use of logistic regression coefficients and SHAP (SHapley Additive exPlanations) values, as they allow transparent identification of clinical features that are critical to making predictions. The results of the experiment show that the suggested framework attains an overall accuracy of approximately 94% on an unseen test set, with strong precision and recall of the minority class, thus proving that the combination of class balancing, cross-validation, and explainable ML results in the outcomes of reliable and clinically credible predictions. All performance results are evaluated on an untouched original test set, while SMOTE is applied strictly within cross-validation folds to prevent data leakage. Unlike many existing studies, the proposed framework ensures leakage-free validation, robust cross-validation, and integrated interpretability for clinically meaningful prediction. Although synthetic sampling improves minority class learning, the model is evaluated carefully to ensure generalization on real-world data. This paper indicates that a rigorously conducted methodology and interpretability in machine learning development are crucial in creating machine learning solutions in healthcare decision support, which is the pathway to real applications in diabetes risk assessment.