Analysis Finds No Widely Used Dataset Documentation Standard for Health AI

Artificial intelligence is transforming healthcare, with much of the progress depending on the quality and documentation of training datasets. Concerns have centered on algorithmic biases that can arise from those datasets. The study compared five standardized methods: Datasheet, Dataset Nutrition Label, Accountability Documentation, Healthsheet, and Data Card.

Researchers evaluated how well each aligned with the STANDING Together Recommendations for Documentation of Health Datasets. They also reviewed real-world usage patterns and collected input from people who generate and use health datasets. The analysis concluded that none of the five approaches are used widely.

It further found that none are fully suited for health datasets. The authors, including researchers from multiple universities and institutions, recommended creation of a standard documentation approach specifically for health datasets.

The paper called for clear guidelines to accompany any new standard. It also urged development of automation tools to support broader adoption. The work was supported by the US National Institutes of Health through grant OT2OD032644. The corresponding author is affiliated with the California Medical Innovations Institute in San Diego.

The article was received on November 4, 2025, accepted on April 25, 2026, and published on May 9, 2026.

“We recommend developing a standard documentation approach for health datasets along with clear guidelines and automation tools to support adoption.”