Run-Level Comparison of Binary Swarm Optimizers for Gallstone Disease Feature Selection

Jabesh Nehemiah Wijaya (Universitas Surabaya, Surabaya)

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 7 No 3 (2026): August 2026

Gallstone disease prediction from non-imaging clinical variables can support early risk assessment, but redundant measurements may reduce model efficiency and complicate interpretation. This study compares five binary swarm optimizers for wrapper feature selection on the UCI Gallstone dataset: Golden Jackal Optimization (bGJO), Egret Swarm Optimization (bESOA), African Vultures Optimization (bAVOA), Tuna Swarm Optimization (bTSO), and FOX Optimization (bFOX). The main contribution of this study is a reproducible run-level statistical comparison of five binary swarm optimizers under an identical wrapper evaluation framework, providing evidence on their accuracy–parsimony and computational trade-offs for gallstone feature selection. The dataset contains 319 complete records, 38 predictors, and an approximately balanced binary outcome. Candidate subsets were evaluated with standardized five-nearest-neighbor classification and stratified 10-fold cross-validation. A weighted objective combined classification error (0.9) and selected-feature proportion (0.1). Each optimizer used 100 epochs, a population of 10, and 10 independent seeded runs. Run-level comparisons used Kruskal–Wallis tests followed by Holm-adjusted Mann–Whitney tests. bGJO achieved the highest mean accuracy (77.56%) with 13.0 features, whereas bAVOA obtained the best mean fitness (0.2270) with only 3.7 features. bTSO selected the fewest features (2.8) but showed greater variability, and bFOX produced the lowest accuracy (67.68%). Accuracy, fitness, feature count, runtime, and memory differed globally among algorithms (all p<0.001). The findings demonstrate a clear accuracy–parsimony trade-off: bGJO is preferable when predictive accuracy dominates, while bAVOA offers the strongest overall compromise under the specified objective. External validation and stability analysis of selected variables are required before clinical use.