Facilitating Clinical Information Extraction with Synthetic Data and Ontology using Large Language Models

Area of Science:

• Natural Language Processing
• Artificial Intelligence in Healthcare

Background:

• Electronic health records contain vast unstructured clinical text.
• Developing information extraction systems is crucial but limited by scarce annotated data.

Purpose of the Study:

• To explore large language models for generating synthetic clinical data for named entity recognition.
• To assess the impact of synthetic data on model performance and generalizability.

Main Methods:

• A novel framework using self-verified synthetic data generation with SNOMED-CT semantic mapping.
• Leveraging GPT-4o-mini for data creation and LLaMA-3-8B for fine-tuning.
• Iterative verification and anomaly detection to refine synthetic data quality.

Main Results:

• Self-verification and semantic mapping significantly improve synthetic data utility.
• A 1:1 ratio of human-annotated to synthetic data yielded optimal performance gains.
• Improved model generalizability observed across four diverse clinical datasets.

Conclusions:

• Synthetic data generation is a scalable solution for clinical NLP annotation challenges.
• Balancing human and synthetic data is key to enhancing model performance.
• The proposed framework advances clinical information extraction capabilities.