Related Experiment Videos
A simple error classification system for understanding sources of error in automatic speech recognition and human
Atif Zafar1, Burke Mamlin, Susan Perkins
1School of Medicine, Regenstrief Institute, Indiana University, 1001 West 10th Street, RG5 Indianapolis, IN 46202, USA. azafar@iupui.edu
International Journal of Medical Informatics
|August 25, 2004
Summary
This study identified nine common error categories in clinical notes from automatic speech recognition (ASR) and human transcription. Developing classification rules helps improve transcription accuracy and clinical documentation quality.
Area of Science:
- Medical Informatics
- Clinical Documentation
- Natural Language Processing
Background:
- Clinical notes are essential for patient care and communication.
- Accuracy in clinical documentation is critical for patient safety and effective healthcare delivery.
- Both automatic speech recognition (ASR) and human transcription are used for generating clinical notes, each with potential for errors.
Purpose of the Study:
- To identify and categorize common errors in clinical notes generated by ASR and human transcription.
- To develop a systematic classification system for these transcription errors.
- To understand the root causes of errors to inform process improvements and enhance accuracy.
Main Methods:
- Integrated Dragon NaturallySpeaking v4.0 ASR into the Regenstrief Medical Record System.
- Captured uncorrected ASR output and compared with uncorrected human-transcribed notes.
- Domain experts analyzed notes, collaboratively defining and refining nine error categories.
Main Results:
- Identified nine distinct error categories: annunciation, dictionary, suffix, added words, deleted words, homonym, spelling, nonsense, and critical errors.
- Developed a classification system based on these categories for systematic error analysis.
- Demonstrated that errors can be meaningfully categorized for further investigation.
Conclusions:
- A straightforward method for examining and classifying transcription errors in clinical documents was established.
- Error classification aids in identifying sources of inaccuracies in clinical notes.
- Implementing targeted measures, such as improved training and system enhancements, can optimize transcription error rates.