Related Experiment Video
Updated: Jan 17, 2026
06:52
Behavioral and Locomotor Measurements Using an Open Field Activity Monitoring System for Skeletal Muscle Diseases
Published on: September 29, 2014
54.5K
Accelerating Medical Record Data Abstraction and Analysis in Muscular Dystrophy: Large Language Models and
Huixue Zhou1,2, Geetanjali Rajamani3, Jiatan Huang2
1Institute for Health Informatics, University of Minnesota, Minneapolis.
Neurology. Clinical Practice
|September 25, 2025
Summary
Large language models (LLMs) show promise for accelerating electronic medical record (EMR) abstraction in muscular dystrophy research, though current performance lags behind manual methods. International Classification of Diseases (ICD) code counts may aid in identifying high-yield cases.
Area of Science:
- Medical Informatics
- Genetics
- Neurology
Background:
- Muscular dystrophies are progressive muscle disorders requiring efficient data abstraction from electronic medical records (EMRs) for research and surveillance.
- Manual EMR abstraction is time-consuming and labor-intensive, necessitating the exploration of automated methods.
- Large Language Models (LLMs) and International Classification of Diseases (ICD) code meta-analysis are investigated as potential solutions.
Purpose of the Study:
- To evaluate the effectiveness of LLMs and ICD code meta-analysis in accelerating EMR data abstraction for muscular dystrophy cases.
- To compare the performance of LLMs against manual annotation for key clinical features of muscular dystrophy.
- To assess the utility of ICD code counts in predicting diagnostic certainty for Duchenne muscular dystrophy (DMD) and limb-girdle muscular dystrophy (LGMD).
Main Methods:
- A cross-sectional study involving EMRs from 44 patients (22 DMD, 22 LGMD) manually annotated for symptoms, ambulatory status, CK levels, and genetic results.
- Five LLMs were tested on clinic notes from these cases, with performance measured by F1 scores against manual annotations.
- A separate cohort of 77 DMD and 59 LGMD patients was analyzed to correlate ICD code encounter counts with diagnostic certainty.
Main Results:
- Manual annotation achieved high inter-rater agreement (80-100%).
- The best-performing LLM, Llama 3-8b, showed variable accuracy (46.8%-69.2%) across different features, generally lower than manual abstraction.
- A high number of ICD code encounters (≥20 for DMD, ≥25 for LGMD) correlated with definite or probable diagnoses.
Conclusions:
- LLMs offer potential for accelerating EMR abstraction in rare diseases like muscular dystrophy, but currently do not match manual abstraction accuracy for unstructured data.
- Llama 3-8b demonstrated the highest performance among the tested LLMs.
- ICD code metadata, specifically encounter counts, can be a valuable tool for prioritizing cases in surveillance and research efforts.
