Biomedical Data Manifest: A lightweight data documentation mapping to increase transparency for AI/ML.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Science

Background:

• Biomedical machine learning (ML) models require robust dataset documentation for clinical decision-making.
• Current documentation methods are burdensome and conflate data/model accountability, especially for non-ML datasets.
• Fairness in ML algorithms depends on user awareness of data issues like provenance and quality.

Purpose of the Study:

• To address gaps in current ML data documentation practices.
• To develop a practical documentation framework for biomedical datasets.
• To improve transparency and bias mitigation in ML applications.

Main Methods:

• Derived consensus documentation fields by mapping elements across four key templates.
• Surveyed biomedical stakeholders (clinicians, bench scientists, data managers, computationalists) on field importance.
• Developed the Biomedical Data Manifest, a modular template with persona-specific field presentation.

Main Results:

• Identified role-dependent prioritization differences among biomedical stakeholders.
• The Biomedical Data Manifest reduces generator burden by offering tailored information.
• Ensures end-users receive role-relevant data information for better ML application.

Conclusions:

• The Biomedical Data Manifest enhances transparency for datasets in public/controlled repositories.
• Improves bias mitigation in machine learning applications through better data understanding.
• Facilitates responsible data sharing and utilization in biomedical research.