Real-World Medical Mistakes: 100 Clinical Reports with Physician-Inserted Errors

100 de-identified clinical reports from Internal Medicine and Emergency Departments containing physician-inserted medical errors, created by Vrda for the Clinipal project in 2026. This dataset provides a specialized ground truth for evaluating AI systems designed to detect patient safety risks in clinical documentation.

Use Cases

• Training text classification models to identify safety risks in clinical-notes
• Evaluating text generation models for their ability to correct physician-inserted mistakes
• Benchmarking NLP performance on specialized internal-medicine terminology

Strengths

• Physician-inserted realistic errors
• De-identified real-world clinical reports
• Focuses on high-stakes Internal Medicine and Emergency Department contexts

Limitations

• Small sample size of 100 reports
• Errors are synthetic/inserted rather than naturally occurring
• Geographic bias toward US clinical practices

Provenance

Source

Clinipal project by Vrda

Collection Method

Physician-annotated modification of de-identified clinical reports

Freshness

Last updated February 2026.

Geography

United States