Refined Feature Selection Scheme-Based Machine Learning Model With SMOTE And Hyperparameter Tuning Approach For Diabetes Prediction

Abstract

Diabetes mellitus is a syndrome that runs with lifelong metabolism that is affecting numerous individuals and requires high precision in early diagnosis in order to minimize complications and related expenses. This paper is suggested a novel hybrid framework for classification of diabetes that applies Recursive Feature Elimination (RFE), Synthetic Minority Oversampling Technique (SMOTE), hyperparameter tuning and SVM learning. In the initial stages the phase of data preprocessing applied over the Pima Indian Diabetes Dataset (PIDD) is by executing method of mean imputation and min-max normalization. It helped to increase the data information quality. Then, the RFE algorithm is applied for extraction of the significant features from the dataset: Blood Pressure, Glucose, Body Mass Index (BMI) and Insulin. For solving the purpose of the class imbalance issue, the SMOTE technique is applied prior to modeling, whereas the hyperparameter tuning is performed using Grid Search method. Proposed model is compared with several machine learning classifiers, such as Random Forest (RF), Logistic Regression (LR), K-Nearest Neighbor (KNN), Histogram Gradient Boosting (HGB), and Naïve Bayes (NB). Experimental results have shown that the suggested RFE-SMOTE-SVM model with hyperparameter tuning demonstrates the best performance and reaches 93.1% accuracy, 92.5% precision, 92.4% recall, 91.1% F1-score and AUC equal to 0.97. Comparative analysis proves that the suggested framework outperforms existing approaches by resolving feature redundancy, class imbalance and classifier optimization issues.