An Interpretable and Scalable Diabetes Detection Model Leveraging GWO-Optimized Features and SVM with Advanced Data Preprocessing

A new systematic approach for preliminarily detecting diabetes utilizes complex data cleaning methods to support a nature-based feature selection method while achieving high-level classification outcomes. The research method starts with extensive data cleaning that fixes noise and addresses missing information and data inconsistencies before performing outlier detection through the Interquartile Range (IQR) approach with Synthetic Minority Over-sampling Technique (SMOTE) implementation to solve common medical dataset class imbalance issues. The implementation of GWO feature selection operation led to a reduction of model complexity by selecting efficient predictive variables which comprise Insulin and Glucose together with Age and Pregnancies attributes. Support Vector Machine (SVM) classifier receives these strategically chosen features for training purposes leading to outstanding results with 0.9884 accuracy evaluation and excellent precision while maintaining recall and F1-score metrics. Adding biologically inspired optimization methods to robust machine learning technology increases prediction accuracy and improves operational speed and delivers medical relevance to results. This work contributes a scalable and interpretable solution for diabetes risk assessment, with potential applicability in real-world healthcare decision-support systems. Future research will aim to incorporate longitudinal data and ensemble learning strategies to further optimize predictive capabilities across diverse clinical environments.