Using Large Language Models to Extract Core Injury Information From Emergency Department Notes
10.3346/jkms.2024.39.e291
- Author:
Dong Hyun CHOI
1
;
Yoonjic KIM
;
Sae Won CHOI
;
Ki Hong KIM
;
Yeongho CHOI
;
Sang Do SHIN
Author Information
1. Department of Biomedical Engineering, Seoul National University College of Medicine, Seoul, Korea
- Publication Type:Original Article
- From:Journal of Korean Medical Science
2024;39(46):e291-
- CountryRepublic of Korea
- Language:English
-
Abstract:
Background:Injuries pose a significant global health challenge due to their high incidence and mortality rates. Although injury surveillance is essential for prevention, it is resource-intensive.This study aimed to develop and validate locally deployable large language models (LLMs) to extract core injury-related information from Emergency Department (ED) clinical notes.
Methods:We conducted a diagnostic study using retrospectively collected data from January 2014 to December 2020 from two urban academic tertiary hospitals. One served as the derivation cohort and the other as the external test cohort. Adult patients presenting to the ED with injury-related complaints were included. Primary outcomes included classification accuracies for information extraction tasks related to injury mechanism, place of occurrence, activity, intent, and severity. We fine-tuned a single generalizable Llama-2 model and five distinct Bidirectional Encoder Representations from Transformers (BERT) models for each task to extract information from initial ED physician notes. The Llama-2 model was able to perform different tasks by modifying the instruction prompt. Data recorded in injury registries provided the gold standard labels. Model performance was assessed using accuracy and macro-average F1 scores.
Results:The derivation and external test cohorts comprised 36,346 and 32,232 patients, respectively. In the derivation cohort’s test set, the Llama-2 model achieved accuracies (95% confidence intervals) of 0.899 (0.889–0.909) for injury mechanism, 0.774 (0.760–0.789) for place of occurrence, 0.679 (0.665–0.694) for activity, 0.972 (0.967–0.977) for intent, and 0.935 (0.926–0.943) for severity. The Llama-2 model outperformed the BERT models in accuracy and macro-average F1 scores across all tasks in both cohorts. Imposing constraints on the Llama-2 model to avoid uncertain predictions further improved its accuracy.
Conclusion:Locally deployable LLMs, trained to extract core injury-related information from free-text ED clinical notes, demonstrated good performance. Generative LLMs can serve as versatile solutions for various injury-related information extraction tasks.
