A Large Language Model Pipeline for Stroke Staging Using Electronical Medical Records.

OBJECTIVE

Identifying stroke patients at different disease stages is a prerequisite for clinical research using electronic medical records (EMRs), whereas an artificial intelligence-based model that can be directly applied remains lacking. We therefore develop a large language model (LLM) pipeline for stroke staging model (StrokeSM) in retrospective clinical research.

METHODS

StrokeSM was developed using a Chinese national stroke database comprising EMRs from 33,637 patients. A total of 2000 patients were randomly selected from the Tianjin regional stroke database for external validation. StrokeSM comprised three phases: stroke hospitalization identification based on BERT and a bidirectional cross-attention network to fuse present illness history and discharge diagnosis, symptom-time extraction based on chief complaint through a UIE-base LLM, and stroke staging classification according to the predefined rules.

RESULTS

On the test set, StrokeSM achieved accuracy, F1 score, precision, and recall of 0.90, 0.91, 0.91, and 0.90, respectively. The F1 score, precision, and recall of StrokeSM for acute phase was 0.91, 0.89, and 0.93, respectively. On the external validation set, StrokeSM had an accuracy, F1 score, precision, and recall of 0.92, 0.93, 0.93, and 0.92, respectively. Moreover, StrokeSM performed remarkably well in acute phase, with F1 score, precision, and recall of 0.97, 0.98, and 0.96, respectively.

CONCLUSIONS

StrokeSM had achieved state-of-the-art performance, providing an accurate method of classifying stroke populations with different disease stages in EMRs, especially in the acute phase. StrokeSM heralds automatic and accurate identification of disease stage phenotypes based on LLM in EMRs, laying the foundation for drawing reliable conclusions in clinical research.