Related Experiment Videos
Deep learning-based natural language processing for critical care identification in pediatric emergency department.
Jiyoung Agatha Kim1, Sangyeon Cho2, Minyoung Hwang3
1Department of Emergency Medicine, Seoul St. Mary's Hospital, College of Medicine, The Catholic University of Korea, Seoul, Korea.
BMC Emergency Medicine
|June 16, 2026
Summary
Deep learning Natural Language Processing (NLP) models, specifically KM-BERT, show promise in identifying critically ill pediatric patients in emergency departments. This AI approach integrates text and vital signs for more accurate and timely predictions than traditional methods.
Area of Science:
- Artificial Intelligence in Medicine
- Pediatric Emergency Care
- Natural Language Processing
Background:
- Pediatric emergency patient assessment is complex, with limited utility of existing scoring systems due to data unavailability.
- Deep learning (DL)-based Natural Language Processing (NLP) can analyze physician notes for early critical illness prediction.
- This study investigates DL-NLP's ability to identify critical care needs using combined structured and unstructured data in pediatric emergency departments (PEDs).
Purpose of the Study:
- To evaluate the effectiveness of DL-based NLP models in identifying critically ill pediatric patients.
- To compare DL-NLP performance against traditional machine learning (ML) models.
- To assess the integration of structured (vital signs) and unstructured (clinical text) data for improved prediction accuracy.
Main Methods:
- Utilized four ML models (logistic regression, extreme gradient boosting, gradient boosting, random forest) and two DL-NLP models (KM-BERT framework).
- Tested models on 87,748 pediatric patient records from a tertiary hospital (2012-2021).
- Included critical cases (0.7%) and admissions (16.1%) for comprehensive analysis.
Main Results:
- The KM-BERT model, employing Masked Language Model (MLM), achieved the highest Area Under the Receiver Operating Curve (AUROC) of 0.941 ± 0.015 and accuracy of 0.952 ± 0.002 for predicting critically ill patients.
- KM-BERT also outperformed other models in predicting hospitalized patients, with an AUROC of 0.879 ± 0.004 and accuracy of 0.798 ± 0.002.
- DL-NLP models demonstrated superior predictive capabilities compared to conventional ML approaches.
Conclusions:
- DL-based NLP, particularly KM-BERT with MLM, shows potential for enhancing early identification of critical and hospitalized pediatric patients in emergency settings.
- Findings suggest DL-NLP integration could improve upon current ML-based models for pediatric critical care prediction.
- Further prospective, multi-center validation is necessary to confirm clinical utility and integration feasibility.
