Social Determinants Of Health Annotations From EHR Text

SHADR is a dataset for evaluating large language models in extracting Social Determinants of Health from free-text electronic health records. It was created by author m720 and uploaded to Hugging Face in December 2023. The dataset includes human annotations on SDoH information found in clinical narratives.

Use Cases

• Benchmarking LLM performance on extracting specific SDoH factors like housing status or employment from EHR free-text sections.
• Training named entity recognition models to identify mentions of demographic and socioeconomic features within clinical narratives.
• Evaluating the utility of synthetic clinical text, as referenced in the study, for data augmentation in SDoH extraction tasks.
• Analyzing annotation patterns to understand the frequency and context of documented versus undocumented SDoH in patient records.

Strengths

• Dataset is based on a formal study examining LLM capabilities for a specific clinical NLP task.
• Includes human-annotated ground truth data for SDoH extraction from real EHR text.

Limitations

• Dataset size, row count, and column details are unknown, limiting assessment of scale and feature diversity.
• Potential geographic or institutional bias is present as the source EHRs are unspecified.

Provenance

Source

Hugging Face user m720, derived from a study on LLMs and EHRs.

Collection Method

Created from a study examining LLM extraction of SDoH from free-text sections of Electronic Health Records.

Freshness

Last updated on Hugging Face in December 2023.