Development of reports preparation system for colonoscopy using speech-to-text technology

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Endoscopy Technology

Background:

• Traditional colonoscopy reporting is inefficient and error-prone, relying on manual methods.
• Improving endoscopic workflows requires faster and more accurate report generation.
• Real-time documentation during procedures is a key area for enhancement.

Purpose of the Study:

• To develop and validate a Speech-to-Text (STT) system tailored for colonoscopy report preparation.
• To enable real-time documentation during colonoscopy procedures.
• To enhance the efficiency and accuracy of endoscopic reporting.

Main Methods:

• Utilized 12,691 colonoscopy reports and 190 minutes of acoustic data for system development.
• Employed deep neural networks with medical vocabulary and noise-cancellation for acoustic and language processing.
• Evaluated system performance using Word Error Rate (WER) and Word Accuracy (WAcc) with data from qualified endoscopists.

Main Results:

• The developed STT system achieved a Word Accuracy (WAcc) of 94.7% during initial development.
• Iterative training and model adaptation led to significant reductions in transcription errors.
• Validation with endoscopists demonstrated a WAcc exceeding 95%, with individual endoscopist accuracies around 86-88%.

Conclusions:

• The study confirms the feasibility and high accuracy of an STT system for colonoscopy reporting.
• The system shows significant potential for enhancing endoscopic workflow efficiency.
• Further research is needed to broaden data diversity and optimize the system for widespread clinical use.