Related Experiment Video
Updated: Jul 2, 2026
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
A Large Language Model Pipeline for Stroke Staging Using Electronical Medical Records
Xiaochao Luo1,2,3,4,5, Yanmei Liu1,2,3,4,5, Yu Ma1,2,3,4,5
1Clinical Epidemiology and Evidence-Based Medicine Center, West China Hospital, Sichuan University, Chengdu, China.
Journal of Evidence-Based Medicine
|July 1, 2026
Summary
A new AI model, StrokeSM, accurately stages stroke patients using electronic medical records (EMRs). This large language model (LLM) pipeline improves clinical research by enabling precise patient classification, particularly in the acute phase.
Area of Science:
- Medical Informatics
- Artificial Intelligence
- Neurology
Background:
- Accurate staging of stroke patients in electronic medical records (EMRs) is crucial for clinical research.
- Existing artificial intelligence (AI) models lack direct applicability for stroke staging in retrospective studies.
Purpose of the Study:
- To develop a large language model (LLM) pipeline, named StrokeSM, for automated stroke staging in retrospective clinical research using EMRs.
Main Methods:
- StrokeSM was developed using a large Chinese national stroke database (33,637 patients) and validated on 2000 patients.
- The pipeline involves three phases: stroke hospitalization identification (BERT, bidirectional cross-attention), symptom-time extraction (UIE-base LLM), and stroke staging classification.
Main Results:
- StrokeSM achieved high performance on the test set (accuracy 0.90, F1 0.91) and external validation set (accuracy 0.92, F1 0.93).
- The model demonstrated exceptional performance in identifying acute phase stroke patients (F1 0.97).
Conclusions:
- StrokeSM offers a state-of-the-art, accurate method for classifying stroke patients by disease stage in EMRs.
- This LLM-based approach facilitates automatic identification of disease phenotypes, supporting reliable clinical research conclusions.
