ClinStigma-SUD v1.0

ClinStigma-SUD is a clinician-adjudicated synthetic benchmark for evaluating contextual stigmatizing language in substance use disorder clinical documentation.

It is the inaugural implementation of ClinStigma, a clinician-guided framework for developing synthetic clinical-language benchmarks.

Coming Soon

Full dataset release is planned for the week of October 19th, 2026.

The release will include:

• 500 synthetic clinical documentation examples
• final clinician-adjudicated labels
• benchmark generation categories
• annotation rubric
• evaluation prompts
• accompanying documentation

The benchmark includes explicit stigmatizing language, neutral documentation, and deliberately ambiguous "gray-zone" examples designed to require contextual interpretation rather than keyword recognition alone.

Benchmark Overview

ClinStigma-SUD v1.0 contains:

• 500 synthetic clinical sentences
• 251 final stigmatizing reference labels
• 249 final neutral reference labels
• 100 intentionally ambiguous gray-zone examples embedded within the final binary reference classes

Independent clinician review demonstrated 98.8% agreement across the full benchmark (Cohen's κ = 0.976) and 94.0% agreement within the intentionally ambiguous subset (κ = 0.880).

Privacy

All ClinStigma-SUD benchmark examples are synthetic and contain no protected health information.

External evaluation reported in the associated manuscript used restricted MIMIC-IV data and previously released Harrigian annotations available to credentialed users through PhysioNet. Restricted source annotations, reconstructed MIMIC-IV context windows, and derived model outputs are not included in this dataset.

License

ClinStigma-SUD v1.0, the annotation rubric, evaluation prompts, and accompanying documentation are released under CC BY 4.0.

Project Repository

https://github.com/izzomd/ClinStigma

Citation

Citation information will be added following release of the associated preprint.