In a multi-target evaluation of medical diagnostic algorithms, researchers found that dramatic cost reductions previously attributed to cost-sensitive feature selection often stem from dataset artifacts and data leakage rather than genuine clinical signal. Auditing ten distinct diagnostic targets across eleven feature pool configurations revealed that reported predictive performance frequently collapsed under rigorous screening, with only dementia and central auditory processing disorder maintaining their validity.
Cost-sensitive feature selection is widely studied as a method to lower diagnostic expenses in public healthcare by trading small amounts of accuracy for reduced test-acquisition costs, serving as an instrument for value-based procurement. However, real clinical data often contain hidden data leakage and structural dataset artifacts that can artificially inflate the apparent efficiency of these algorithms, vulnerabilities that the authors note are rarely evaluated in the computational intelligence literature.
To determine when cost-sensitive and fuzzy feature selection genuinely aids clinical decision-making, the study analyzed a clinical cohort comprising 1,929 older adults alongside assigned per-test tariffs. The authors developed and tested three feature selection strategies that combined Extreme Gradient Boosting, or XGBoost, with a fully operationalized fuzzy layer. This fuzzy framework incorporated explicit membership functions, an alpha-cut acceptability rule, triangular fuzzy costs, and Mamdani merit inference, with all evaluations conducted within a leakage-free nested cross-validation protocol using standard baseline models and statistical tests.
The central methodological innovation introduced in the study is a pair of reusable diagnostic guards aimed at detecting hidden data flaws. One guard is designed to detect target circularity and determinism, while the other guard specifically identifies monotonicity and discretization artifacts. When applied across the ten diagnostic targets, these screening guards demonstrated that the majority of evaluated conditions carried negligible predictive signal when subjected to independent tests.
Under the audit of these guards, several prominent diagnostic claims were identified as analytical artifacts. Specifically, the study revealed that a previously documented outcome claiming frailty could be diagnosed about seven-fold cheaper while preserving approximately 96 percent accuracy was in fact an artifact of target circularity. In a similar manner, diagnostic predictions for age-related hearing loss were demonstrated to be an artifact of data discretization.
Among all the conditions evaluated, only dementia and central auditory processing disorder survived both testing guards without collapsing. In the case of dementia assessment, the experimental results demonstrated that an inexpensive triage strategy relying solely on patient education and age preserved almost all of the predictive accuracy achieved by an extensive clinical diagnostic battery priced at €413.
Evaluating the fuzzy architecture across 1,000 tariff perturbations showed that the fuzzy layer did not alter the ordering of selected feature subsets. Instead, the fuzzy system delimited the 31.4 percent of diagnostic comparisons that a tariff uncertainty of plus or minus 15 percent renders undecidable, a specific boundary that encompassed every observed subset inversion. Consequently, the researchers concluded that the operationalized fuzzy layer provides graded, tariff-aware decision support rather than gains in predictive accuracy, demonstrating that cost savings are meaningful only when underlying signals withstand leakage and artifact guards.
Subject of Research: Cost-sensitive feature selection in medical diagnostics
Article Title: Cost-Sensitive Feature Selection in Medical Diagnostics: A Leakage- and Artifact-Aware Multi-Target Study with Fuzzy Decision Support
Article References: Franchini, F., Pirlo, G., Sarcinella, L., & Semeraro, G. (2026). Cost-Sensitive Feature Selection in Medical Diagnostics: A Leakage- and Artifact-Aware Multi-Target Study with Fuzzy Decision Support. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01665-6
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01665-6
Keywords: cost-sensitive feature selection, medical diagnostics, data leakage, dataset artifacts, fuzzy decision support, XGBoost, nested cross-validation, target circularity, discretization artifacts, dementia screening, central auditory processing disorder, older adults, Mamdani inference, test-acquisition costs, tariff uncertainty, frailty diagnosis, value-based healthcare
News Source: Denise Maddox. (November 17, 2026). Most Cost Savings in Medical Feature Selection Stem From Dataset Artifacts. Scienmag.



