Enhancing SVM Performance Accuracy for Diabetes Diagnosis Using an Improved Ant Colony Optimization Based Support Vector Machine

Accurate diagnosis of diabetes is crucial for effective management and improved patient outcomes. Traditional Support Vector Machine (SVM) classifiers often struggle with accuracy due to parameter optimization challenges and unbalanced datasets. These challenges were addressed by developing an improved pheromone update technique for Ant Colony Optimization ACO-optimized SVM classifier. To achieve the aforementioned, the research generated a Hybrid Adaptive Pheromone Update Technique (HAPUT), Dynamic Exploration-Exploitation Balance (DEEB) and Pheromone Influence Factor (PIF). Subsequently, the parameters, BoxConstraint and KernelScale of the Support Vector Machine (SVM) classifier were optimized using an Ant Colony Optimization (ACO) approach in which HAPUT was used as the ACO pheromone update technique. Hence, each ant selects SVM parameters based on pheromone levels. The model developed was run in MATLAB codes using the PIMA Indian Dataset (PID) which composed of 268 diabetic and 500 non- diabetic samples. The dataset was split into 80/20 for training and validation. Thus, the accuracy of ACO-optimized SVM for default and improved pheromone update were compared.The comparative analysis shows that SVM has the optimum performance with accuracy, precision and recall of 79.13%, 69.388 % and 50.746%, respectively; while ACO optimized with SVM has the optimal accuracy and precision of 83.0435 % and 80.9524 %. Moreso, the results of the ACO-optimized SVM with a Default Pheromone Update Technique (DPUT) and ACO-optimized SVM with an Improved Pheromone Update Technique (IPUT) shows that IPUT reflected higher performance of 86.520 %, 81.130 % and 67.187 % for accuracy, precision and recall, respectively. This outcome is still optimal when compared to results from related studies. In conclusion, the model developed converges to the best combination of SVM parameters, BoxConstraint (C) and KernelScale, which yields the highest classification accuracy.