Cost-Sensitive Feature Selection in Medical Diagnostics: A Leakage- and Artifact-Aware Multi-Target Study with Fuzzy Decision Support

Cost-sensitive feature selection promises cheaper diagnosis by trading accuracy for savings in test-acquisition cost, a natural instrument for value-based procurement in public healthcare. On real clinical data, however, such promises are easily inflated by data leakage and dataset artifacts, rarely tested for in the literature. This paper therefore asks when cost-sensitive, and specifically fuzzy, feature selection actually helps. On a cohort of 1929 older adults with per-test tariffs, three selection strategies combine Extreme Gradient Boosting (XGBoost) with a fully operationalized fuzzy layer, comprising explicit membership functions, an \(\alpha\) -cut acceptability rule, triangular fuzzy costs and Mamdani merit inference, all evaluated inside a leakage-free nested cross-validation with standard baselines and statistical tests. The central methodological contribution is a pair of reusable guards, one for target circularity and determinism, one for monotonicity and discretization artifacts. Across ten targets in eleven pool configurations, the guards show that the “about seven-fold cheaper at \({\approx }96\%\) accuracy” frailty result is an artifact of target circularity, that age-related hearing loss is a discretization artifact, and that most targets carry little signal in independent tests. Only dementia and central auditory processing disorder survive both guards; for dementia, a near-free education-and-age triage retains almost all the predictability of a €413 battery. Over 1000 tariff perturbations the fuzzy layer does not reorder subsets but delimits the \(31.4\%\) of comparisons that \(\pm 15\%\) tariff uncertainty leaves undecidable, a band containing every observed inversion. It thus contributes graded, tariff-aware decisions rather than accuracy: savings are real only where the signal survives both guards.