Feasibility study of speech recognition for gathering information needs.
Karthik Natarajan1, Robert F Duffy, Stephen B Johnson
1Department of Biomedical Informatics, Columbia University, New York, NY, USA.
AMIA ... Annual Symposium Proceedings. AMIA Symposium
|August 13, 2008
Summary
Automated speech recognition (ASR) in medicine needs improvement. Even with basic training, ASR struggles to accurately capture clinician information needs in noisy settings, impacting medical data quality.
Area of Science:
- Medical Informatics
- Speech Technology
- Clinical Documentation
Background:
- Automated speech recognition (ASR) is increasingly utilized in healthcare.
- Limited research exists on ASR's efficacy in capturing clinician information needs within noisy clinical environments.
Purpose of the Study:
- To evaluate the performance of ASR in transcribing clinician-generated questions in real-world noisy settings.
- To assess the semantic and syntactic accuracy of ASR-generated medical transcriptions.
Main Methods:
- Analysis of 72 clinician-generated questions transcribed by ASR.
- Evaluation of transcriptions for semantic and syntactic errors.
- Assessment of ASR performance in capturing nuanced clinical information.
Main Results:
- Significant semantic and syntactic errors were identified in ASR transcriptions.
- Basic user training proved insufficient for accurate semantic capture in noisy environments.
- ASR's ability to reliably record clinician information needs was compromised.
Conclusions:
- Current ASR systems require further development for reliable use in noisy medical settings.
- Enhanced training protocols or advanced ASR algorithms are necessary to improve accuracy.
- The fidelity of clinician information capture via ASR in challenging acoustic conditions remains a concern.
