Exploration and Practice of Extracting Structured Medical Information from Clinical Triage Records Using Large Language Models

This article introduces an open-source project that explores how to use large language models (LLMs) combined with fine-tuning, prompt engineering, and traditional biomedical NLP techniques to extract structured medical data from unstructured clinical triage records, aiming to provide solutions for clinical decision support and medical data standardization.

In modern healthcare systems, a large amount of clinical information (such as triage records and discharge summaries) exists as unstructured text. Traditional methods struggle to efficiently extract key information; manual sorting is time-consuming, labor-intensive, and prone to omissions. Standardization and structuring of medical data are core challenges in clinical informatics, and the development of LLM technology provides possibilities for automated conversion.

The project adopts a multi-technology integration strategy: 1. Fine-tuning technology: Perform domain-specific fine-tuning of pre-trained models based on the Hugging Face ecosystem to adapt to the medical terminology system; 2. Prompt engineering: Use methods such as few-shot learning and chain-of-thought to quickly adapt to tasks without a large amount of labeled data; 3. Biomedical NLP integration: Combine traditional technologies like NER to improve extraction accuracy and interpretability.

The system mainly extracts three types of key information from clinical texts: 1. Disease diagnosis information: Identify disease names (e.g., Type 2 diabetes, hypertension) and their different expressions; 2. Symptom descriptions: Extract patient symptoms (e.g., acute dyspnea, persistent chest pain); 3. Treatment plans and medications: Extract treatment plans and drug information (e.g., metformin 500mg twice daily).

The project compares two technologies: Prompt engineering has the advantages of no need to modify model parameters, low cost, and fast deployment, making it suitable for rapid verification and task change scenarios, but it depends on the quality of prompts; Parameter-Efficient Fine-Tuning (PEFT) deeply learns domain knowledge through minor parameter adjustments, with more stable performance, but requires more computing resources and labeled data; Experiments provide support for technology selection.

The project has evolved from an experimental sandbox to a modular and reusable pipeline. The plan includes: Refactoring the code into Python scripts to improve maintainability; Expanding the model backend (integrating OpenAI, Anthropic APIs, local deployment of Llama3, etc.); Establishing a standardized evaluation system (F1 score, precision, recall rate, etc.).

The project aims to build an end-to-end reproducible pipeline (input unstructured text → output structured JSON/CSV data). Its clinical value includes: Assisting clinical decision-making, promoting medical data standardization, and accelerating medical research; In the future, with the progress of LLMs and data openness, it will become an important infrastructure for smart healthcare.