<![CDATA[Subscription-gated IEEE DataPort datasets are difficult to verify before reuse because independent inspection requires access most researchers do not have; we obtained direct access to one such case. We apply a six-stage, three-layer audit (provenance, structural/duplication plausibility, target validity) to “An Anonymized Feature-Based Dataset for Lung Cancer Prediction Using Explainable AI Models,” a 15-feature, binary-symptom-checklist dataset presented as newly compiled to support explainable-AI research. The delivered file contains 2,457 rows, of which 2,132 (86.8%) are exact duplicates of another row, leaving 325 unique records. The dataset’s own hosting page links a peer-reviewed “Research Article” co-authored by two of this dataset’s three listed authors; that 2024 publication states its dataframe “contains 309 entries” with “no null values” — figures that match neither the delivered file’s row count nor its 325 unique records. The same publication’s introduction and methods describe a multi-source study drawing on major imaging and genomic repositories, none of which appear in the delivered file or in the publication’s own reported results, which instead use the same 16-column schema and encoding as a long-circulating public dataset. Within the 325 unique records, duplication is class-dependent: unique cancer-negative rows are repeated 14.1 times on average (up to 86), versus 4.5 times for cancer-positive rows, shifting the delivered file’s class balance relative to the unique data without documentation. Separately, the linked publication reports identical accuracy, precision, recall, F1, and AUC-ROC for five structurally different classifiers, a pattern inconsistent with independent model evaluation.]]>
Copyrights © 2025
