Enhanced Diabetes Prediction Through Systematic Analysis of Machine Learning Techniques and Data Optimization

This research concentrates on applying various machine learning algorithms to develop an effective and efficient predictive model for screening for diabetes prediction. Its primary aim is to enhance the performance of the model and time to learn and generalize by performing feature engineering, feature selection, and hyperparameter tuning all apt. The strategy begins with preprocessing data, where problems of missing or incorrect values are addressed and duplicate records are eliminated on a dataset from Kaggle. SMOTE is employed in the case of class imbalance; new samples for the minority class are created, while outliers in datasets are identified using the DBSCAN algorithm. The required features are achieved through the Wolf Optimized Recursive Elimination (WORE), the union of the Grey Wolf Optimization (GWO) algorithm, and Recursive Feature Elimination (RFE). The model is constructed through the Random Forest-Enhanced Logistic Regression (RFELR) model, a hybrid of both Random Forest and Logistic Regression in which Random Forest is employed to obtain a nonlinear relationship, combined with logistic regression for a linear decision boundary. The final adjustment of the models is carried out using Grid Search as the Hyperparameter tuning method. These indicate that the overall performance reaches 91.5 % accuracy while precision, recall, and Fl were at 89.6%, 83.1 %, and 87.5% respectively. The paper has pointed out the appropriateness of the suggested approach in predictive performance improvement for diabetes detection to a significant degree beneficial to clinical use.