Arthroplasty Clinical Notes for NLP Data Extraction, 2012-2018

1,000 unstructured clinical notes from eight surgeons were used to test an NLP algorithm for extracting 38 structured variables related to arthroplasty. The algorithm achieved 96.3% overall accuracy, with performance varying based on data format, such as 98% accuracy for templated data like implant brand. This dataset supports research into automating registry data collection from free-text medical records.

Use Cases

• Training NLP models to extract structured clinical variables from unstructured notes based on the 38 preoperative, operative, and postoperative variables mentioned.
• Benchmarking automated data extraction accuracy for medical registry participation, as described in the study's methodology.
• Comparing NLP performance on different data formats, such as templated data (e.g., range of movement) versus free-text descriptions (e.g., presence of DVT).

Strengths

• The underlying NLP algorithm was validated on a random test dataset, achieving 96.3% accuracy, 95.2% sensitivity, and 97.4% specificity.
• Data covers 1,000 clinical notes from patients undergoing primary arthroplasty between 2012 and 2018, providing a multi-year clinical scope.
• The study design includes a clear comparison between manual and automated extraction for 38 specific clinical variables.

Limitations

• The actual dataset (clinical notes) is not directly available; the description summarizes a research paper's findings.
• Row count and file formats for the underlying notes are unknown, which limits suitability assessment.
• Data may reflect bias inherent to the notes from eight specific surgeons at a single institution.

Provenance

Source

University of California, San Francisco

Collection Method

Clinical and hospital notes were manually collected and variables were extracted for NLP algorithm training and testing.

Time Range

2012-2018