This contribution is about a novel swarm search algorithm designed to enhance the feature selection (FS) performance in building an incremental Bayesian network (IBN) for efficient cancer data classification. IBN has the advantage of finding out casualty relation between the variable and offering reasonable prediction accuracy, both are in demand for medical data analysis (MDA). FS is challenging in this scenario because the choice of the feature subset would have to satisfy both objectives of accuracy and generalization for the IBN. An extensive simulation experiment was carried comparing one dozen of swarm-based FS methods by different contemporary metaheuristics. In particular, a new search method called Teng-Yue Algorithm (TYA) which means motions of "gallop" and "leap" is formulated that mimic no living creation but just the two fundamental dynamics of local intensification and global exploration in the very original metaheuristic search design. It was observed from the experimental results that TYA outperformed all other metaheuristics by pairing with correlation-based feature subset selection and IBN, with a careful tuning on the balance of the local and global searching efforts distribution. The result is encouraging as TYA is relatively simple and able to empower fast and accuracy BN inference that MDA may leverage.