LLM-Driven Lab Result Extraction from Electronic Health Records

Mehmet F Bagci, Yusuf Ozturk, Truong Nguyen, Brian D. Modena, Toan Do, Samantha R. Spierling Bagsic, Anna L. Ritko

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Electronic Health Records, Natural Language Processing, Large Language Models, Structured Data Extraction

TL;DR: We trained and tested multiple large language models for extracting structured spirometry data from unstructured clinical notes. The best-performing model achieved high accuracy

Abstract: Extracting structured lab results from electronic health records (EHRs) is critical for large-scale clinical research and patient care but remains challenging due to the unstructured nature of EHR data. This study evaluates large language models (LLMs)—including Phi4, LLaMa-3, and Qwen2.5—across zero-shot, one-shot, few-shot, and fully fine-tuned settings to assess their ability to extract key lab values from clinical text. We analyze prompt design, adaptation strategies, and the effects of quantization on speed and accuracy. LLaMa-3 (8B) with full fine-tuning achieved the highest accuracy (93.79%), while LLaMa-3.3 (70B) excelled in few-shot mode (89.56%). Incorporating additional spirometry data further improved asthma severity classification accuracy from 0.72 to 0.85. These findings demonstrate that well-designed prompts and fine-tuning enable efficient and accurate LLM-based data extraction, with clear trade-offs between model size, speed, and performance.

Track: 6. Theoretical Biomedical Informatics

Registration Id: LDNXKQD2SVW

Submission Number: 236