<![CDATA[Diabetes mellitus is one of the fastest-growing chronic diseases globally, with the International Diabetes Federation (IDF) reporting approximately 589 million adults affected in 2024 and projecting a rise to 853 million by 2050. Late diagnosis remains a critical barrier to effective management, motivating the development of accurate automated prediction systems. This study proposes a hybrid approach combining Ant Colony Optimization (ACO) as a wrapper-based feature selection method and Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel as the classifier to predict diabetes risk. The Pima Indians Diabetes Dataset from the UCI Machine Learning Repository, comprising 768 samples with 8 clinical features and a binary class label (diabetic/non-diabetic), was used as the experimental benchmark. ACO was employed to identify the optimal feature subset by iteratively updating pheromone trails based on classification fitness, effectively reducing dimensionality while preserving discriminative information. The SVM model was subsequently trained on the five selected features to generate predictions. Experimental results demonstrate that the proposed ACO-SVM hybrid achieves an accuracy of 82.47%, precision of 79.31%, recall of 76.19%, and an F1-Score of 77.72%, outperforming standalone SVM (78.57%), PSO-SVM (80.52%), GA-SVM (79.87%), Random Forest (80.52%), Naïve Bayes (76.30%), KNN (74.03%), and Decision Tree (73.38%). These findings confirm that integrating colony-based metaheuristic optimization with SVM substantially improves diabetes prediction performance. Future work may explore adaptive ACO parameter tuning, larger and more diverse clinical datasets, and ensemble-learning integration to further enhance model robustness.]]>
