Back to all news
Concerns Raised Over Unverified Datasets in AI Health Prediction Models
Tags:Research
A new study finds widely used AI health prediction models are built on datasets with unverifiable origins, raising safety and validity concerns.
Key Details
- 1Researchers at QUT and AusHSI examined two popular health datasets on Kaggle used for stroke and diabetes prediction.
- 2These datasets have been cited in 125 peer-reviewed studies, with little to no information about data provenance.
- 3Three AI models trained on these datasets have been used in clinical practice; one was cited in a medical device patent.
- 4The datasets scored 0/9 on the TRIPOD+AI data provenance criteria, indicating they are unsuitable for clinical use.
- 5Seven articles based on these datasets have already been retracted as unreliable.
- 6Researchers urge journals, funders, and data repositories to enforce stricter data-source disclosure, and recommend removal of the problematic datasets.
Why It Matters
Reliable, transparent datasets are fundamental to developing trustworthy AI in radiology and medicine. Use of unverifiable data directly threatens patient safety and undermines the clinical credibility of predictive AI tools, highlighting the urgent need for rigorous data oversight.
Source
EurekAlert
Related News
•EurekAlert
AI Unlocks Hidden Brain Lesions in Multiple Sclerosis via MRI
AI-enabled image processing reveals invisible cortical lesions in MRI scans of MS patients, advancing diagnosis and research.
•EurekAlert
AI-Driven Molecular Profiling Enhances Thrombosis Risk Prediction
Researchers create an AI model integrating genetic, clinical, and transcriptomic data to identify individuals at risk for thrombosis.
•EurekAlert
AI Pathology Analysis Predicts Immunotherapy Response in Rare Tumors
AI-based analysis of tumor pathology slides can predict immunotherapy outcomes in rare cancers, according to a recent MD Anderson study.