A new analysis warns that dozens of artificial-intelligence models built to predict stroke and diabetes risk were trained on datasets with dubious provenance and implausible patterns—and some may already have reached patients. Researchers led by Queensland University of Technology statistician Adrian Barnett flagged two widely downloaded Kaggle datasets that show signs of fabrication, including near-perfect completeness and unrealistically discrete blood-glucose values. At least two models appear to have been used in hospitals in Indonesia and Spain, with others cited in a U.S. clinic, patent filings and public web tools. The findings—posted as a medRxiv preprint—have prompted journal investigations and fresh calls for strict disclosure of data sources for clinical AI. Critics say tools trained on unverifiable data risk misdiagnosis, inappropriate treatment and reputational damage for publishers and funders. Kaggle declined to say whether it would act on the datasets. The episode underscores rising regulatory and governance pressure on medical AI to meet basic standards of transparency, validation and real-world reliability.
Related articles:
— Ethics and governance of artificial intelligence for health
— AI Risk Management Framework (NIST RMF 1.0)
— TRIPOD: Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis
— Datasheets for Datasets
