The stability of computational intelligence based subset feature selection (CI-SFS) has not been explored. In this study, 44 methods are evaluated on BCDR-F03 using 5 stability estimators. Experimental results identify 3 methods achieving 0.55 or higher scores from two estimators, 7 methods leading to good classification (area under the curve ≥ 0.80) and 4 potential signatures helping cancer diagnosis. Conclusively, most of the CI-SFS methods seem sensitive to data perturbation and different estimators cause inconsistent results. In future work, attention should be paid to developing robust fitness functions to enhance feature preference and designing advanced estimators to quantify the feature selection stability.