Multispecialty Dental EMRs from Chairside Audio: An Exploratory Study

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing
• Clinical Informatics

Background:

• Generating standardized electronic medical records (EMRs) from dental consultations is challenging due to acoustic interference and diverse clinical specialties.
• Existing methods may struggle with accuracy and consistency in real-world dental settings.

Purpose of the Study:

• To develop and evaluate a modular large language model (LLM) system for creating standardized EMRs from dental chairside audio.
• To enhance EMR generation by incorporating evidence-based modules and consensus mechanisms.

Main Methods:

• A pipeline integrating audio enhancement, automatic speech recognition, and a cascaded LLM was developed.
• An evidence-enhanced system (System 2) with multisource evidence capture and consensus voting was compared to a baseline system (System 1).
• 100 deidentified dental outpatient recordings from periodontics, orthodontics, and prosthodontics were used for evaluation.

Main Results:

• System 2 demonstrated improved robustness, significantly reducing variability in EMR generation across dental specialties (e.g., 55.9% reduction in orthodontics).
• The AI system achieved high correlation with human expert ratings for output quality (r=0.823).
• The system showed methodological feasibility for generating accurate and consistent EMRs from authentic clinical audio.

Conclusions:

• The developed LLM system is a feasible tool for generating standardized EMRs from challenging dental audio recordings.
• Evidence-grounding and consensus mechanisms enhance the medical accuracy and consistency of AI-generated EMRs.
• This approach offers a promising solution for reducing variability in complex clinical documentation workflows.