DataAtlas: automatic generation of data dictionaries using large language models

Area of Science:

• Data Science
• Bioinformatics
• Clinical Informatics

Background:

• Dataset reuse is hampered by poor documentation, limiting secondary analysis.
• Automated data dictionaries are needed to improve data interpretability and accessibility.

Purpose of the Study:

• Develop DataAtlas, an open-source system for automated data dictionary generation from tabular datasets.
• Enhance the accessibility, reproducibility, and reuse of clinical data through improved documentation.

Main Methods:

• DataAtlas integrates structural profiling and large language model (LLM)-based semantic inference.
• It generates descriptions using column metadata, statistical summaries, and sample values.
• The system was evaluated on clinical datasets using validation, expert review, and task performance.

Main Results:

• Generated descriptions were often preferred over official documentation, especially when existing ones were incomplete.
• Human expert review confirmed high accuracy and low hallucination rates for LLM-generated descriptions.
• Augmenting database schemas with generated data dictionaries significantly improved text-to-SQL execution accuracy (0.52 to 0.88).

Conclusions:

• Automated data dictionary generation enhances dataset interpretability and downstream analytical performance.
• Column-level metadata, particularly sample values, is crucial for grounding LLM descriptions.
• DataAtlas offers a practical solution for generating structured data dictionaries, boosting clinical data reuse.