Abstract
Abstract
WrapEvoFS provides a deterministic, auditable rule for choosing one feature set from repeated stochastic searches. It retains candidates within a prespecified development-score gap, selects the most representative set by feature overlap, and records the decision before held-out evaluation. A corrected genetic objective preserves score ordering lost through zero truncation. In a fully nested four-class cancer analysis, all 30 locked signatures met the 0.01 tolerance, and 11 differed from the highest-scoring candidate; every changed selection increased mean feature overlap within the eligible pool. Paired differences in area under the receiver operating characteristic curve against recursive feature elimination with cross-validation all had 95% intervals crossing zero. No selector dominated matched comparisons on feature count and predictive performance, and one-time evaluation of frozen signatures at held-out centers yielded heterogeneous effects. WrapEvoFS standardizes selection within a fixed candidate bank under an empirical score-gap constraint, enabling reproducible reporting of performance, compression, and feature-agreement trade-offs without implying predictive superiority or stability under changes in participant composition.