Enhancing Machine Learning Models for Early Detection of Diabetes Using Explainable AI

Understanding and detecting diabetes mellitus early is essential to prevent serious complications such as cardiovascular disease, chronic kidney disease, and neuropathy. Traditional methods of diagnosis including fasting plasma glucose (FPG) and HbA1c tests often require a blood draw and typically diagnose diabetes at an advanced stage. This paper provides an end-to-end investigation involving seven machine learning algorithms for the purpose of diabetes prediction: Logistic Regression, Support Vector Machines (SVMs), KNNs, Decision Trees, AdaBoost, Naive Bayes, and XGBoost. Advanced data preprocessing techniques, including feature scaling, outlier handling, and Synthetic Minority Over Sampling Technique (SMOTE), helped to enhance models. The dataset consisted of 768 patient samples, along with 9 clinical attributes, was pre-processed and evaluated with stratified k-fold cross-validation. Of the explored models, KNN achieved the best prediction accuracy of 87% after tuning hyperparameters, followed closely by SVM with 85% accuracy. An additional step of feature engineering produced new categorical variables for Weight_Status, Age_Category, and Risk_Level to aid in model interpretability. Model explainability was produced with SHapley Additive exPlanations (SHAP) values. The most important predictor feature was glucose (35% importance). A Flask web app was created for real-time risk assessment of diabetes, demonstrating a practical deployment option. The findings suggest that optimized machine learning models can function as valuable instruments for early diabetes detection, potentially enhancing patient prognoses through prompt intervention.