AI-Driven Dental Procedure Coding: A Multi-Model Framework for CDT Extraction from Clinical Text.

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing (NLP)
• Dental Informatics

Background:

• Dental procedure coding is crucial for billing and documentation but is manual and error-prone.
• Automated coding using NLP has advanced medical fields, but dental applications, specifically Code on Dental Procedures and Nomenclature (CDT) extraction, are underexplored.
• Developing AI for automated CDT code extraction from clinical notes is needed.

Purpose of the Study:

• To develop and evaluate an AI framework for automating CDT code extraction from dental clinical notes.
• To compare the performance of large language models (LLMs) and traditional deep learning methods for this task.
• To assess strategies including zero-shot prompting, fine-tuning (QLoRA, PEFT LoRA), and a Bidirectional GRU (Bi-GRU) classifier.

Main Methods:

• Evaluated three LLM strategies: zero-shot prompting, QLoRA fine-tuning, and parameter-efficient fine-tuning (PEFT) with LoRA.
• Utilized a supervised Bidirectional GRU (Bi-GRU) classifier as a baseline.
• Employed a synthetic dataset, structured JSON output schemas, few-shot prompting, and batch inference for evaluation.
• Assessed performance using micro/macro F1 scores, precision, recall, exact-match accuracy, and Hamming loss.

Main Results:

• Zero-shot LLM achieved high micro-F1 (0.9614) but struggled with rare codes and produced hallucinations.
• Fine-tuned LLMs improved domain alignment but over-generated incorrect codes.
• The Bi-GRU model offered balanced performance (micro-F1=0.9362, macro-F1=0.9377) with fewer hallucinations but missed some context-dependent codes.
• PEFT LoRA outperformed QLoRA among fine-tuned LLMs.

Conclusions:

• LLM-based and supervised approaches have complementary strengths for dental coding.
• LLMs offer strong contextual understanding and rapid deployment.
• Traditional models provide stable, precise multi-label classification.
• Hybrid, schema-constrained systems are recommended for scalable dental procedure coding.