Automatically Detecting Likely Edits in Clinical Notes Created Using Automatic Speech Recognition.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Documentation

Background:

• Automatic Speech Recognition (ASR) offers cost reduction for clinical note creation in electronic medical records.
• Current ASR accuracy necessitates manual post-editing by practitioners to maintain note quality.
• Reducing the time spent on editing ASR transcripts is crucial for clinical workflow efficiency.

Purpose of the Study:

• To investigate novel methods for automatic detection of edit regions in ASR-generated clinical notes.
• To identify both potential ASR errors and areas requiring cleanup or rephrasing.
• To reduce the manual editing burden on healthcare practitioners.

Main Methods:

• Development of detection models using logistic regression and conditional random field approaches.
• Exploration of text-based features incorporating clinical note structure and medical context.
• Utilization of diverse medical text resources to enhance feature extraction.

Main Results:

• Experimental validation on a large corpus of practitioner-edited clinical notes.
• Successful detection of 67% of sentence-level edits.
• Successful detection of 45% of word-level edits with a 15% false detection rate.

Conclusions:

• The developed methods show significant potential for automating the identification of necessary edits in ASR clinical notes.
• This approach can substantially decrease the time practitioners spend on post-editing ASR transcripts.
• Automated edit region detection is a promising strategy for improving the efficiency and quality of ASR-assisted clinical documentation.