Optimizing Diabetes Prediction: A Systematic Study of Machine Learning Techniques and Data Enhancements

This study focuses on the application of different machine learning algorithms to derive an efficient and effective predictive model for screening for diabetes prediction. The main objective is to improve the model performance and time to learn and generalize by doing features engineering, feature selection, and hyperparameter tuning all appropriate. The approach starts with data pre-processing, where issues such as missing or wrong values are solved and duplicate records are removed on a given dataset from Kaggle. The ADSYAN algorithm is used for class imbalance; new samples from the minority class are synthesized, while the DBSCAN algorithm is used to detect outliers in datasets. The necessary features are obtained using the Grey Wolf Recursive Elimination (GWRE), the combination of the Grey Wolf Optimization (GWO) technique and Recursive Feature Elimination (RFE). The model is built through the Gradient-Enhanced Logistic Regression (GELR) model, which is a hybrid of both Gradient boosting and Logistic Regression where gradient boosting is used to derive a non-linear relationship, integrated with logistic regression for a linear decision boundary. A final tuning of the models is done using Grid Search as the Hyperparameter optimization technique. These show that the overall performance achieves 91.8 % accuracy while the precision, recall, and F1 values were at 89.6%, 82.1%, and 87.5% respectively. This paper has highlighted the suitability of the proposed methodology in enhancing predictive performance for diabetes detection to a considerable extent valuable to clinical application.