Scholarlink

Background <p>Emergency department (ED) crowding and prolonged boarding times remain major challenges in acute care. We aimed to develop and validate machine-learning (ML) models to predict ED disposition (hospital admission vs. discharge) before physician evaluation using routinely available electronic health record (EHR) data and features extracted from triage notes with large language models (LLMs), to support proactive inpatient bed management and staffing preparedness during high-demand and surge conditions.</p> Methods <p>This retrospective study analyzed 998,109 encounters from 11 emergency departments within a regional health system in South Carolina (January 2023–November 2024). Predictors available before physician evaluation included demographics, arrival characteristics, vital signs, workflow variables, and LLM-extracted features. Nine classifiers were trained with and without class balancing to address the lower admission rate and with and without LLM-derived features. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), recall, precision, F1-score, and accuracy, and Shapley Additive Explanations (SHAP) analysis identified influential predictors.</p> Results <p>Balanced models outperformed unbalanced versions. The top-performing models (CatBoost, XGBoost, LightGBM) achieved AUROC of 0.89, recall &gt; 0.83, and accuracy of 0.79. Incorporating LLM-extracted features further improved performance, primarily by increasing recall and elevating features such as “referral by another clinician” among top predictors. Other key predictors included triage acuity, presenting hospital, chief complaint, age, and arrival mode.</p> Conclusion <p>Integrating LLM-extracted variables with structured EHR data enables accurate early prediction of ED disposition, providing hospital decision-makers with early notice of inpatient demand to support proactive bed allocation, staffing coordination, and surge management.</p> Trial registration <p>Not applicable.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting emergency department disposition using machine learning and large language models to support proactive capacity management: a multicenter retrospective study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月12日

查看全文
  • Farzad Zeinali,
  • Kevin Taaffe,
  • Chris Gaafary,
  • William Jackson,
  • Michael Ramsay,
  • Jessica Hobbs,
  • Ronald Pirrallo

摘要

Background

Emergency department (ED) crowding and prolonged boarding times remain major challenges in acute care. We aimed to develop and validate machine-learning (ML) models to predict ED disposition (hospital admission vs. discharge) before physician evaluation using routinely available electronic health record (EHR) data and features extracted from triage notes with large language models (LLMs), to support proactive inpatient bed management and staffing preparedness during high-demand and surge conditions.

Methods

This retrospective study analyzed 998,109 encounters from 11 emergency departments within a regional health system in South Carolina (January 2023–November 2024). Predictors available before physician evaluation included demographics, arrival characteristics, vital signs, workflow variables, and LLM-extracted features. Nine classifiers were trained with and without class balancing to address the lower admission rate and with and without LLM-derived features. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), recall, precision, F1-score, and accuracy, and Shapley Additive Explanations (SHAP) analysis identified influential predictors.

Results

Balanced models outperformed unbalanced versions. The top-performing models (CatBoost, XGBoost, LightGBM) achieved AUROC of 0.89, recall > 0.83, and accuracy of 0.79. Incorporating LLM-extracted features further improved performance, primarily by increasing recall and elevating features such as “referral by another clinician” among top predictors. Other key predictors included triage acuity, presenting hospital, chief complaint, age, and arrival mode.

Conclusion

Integrating LLM-extracted variables with structured EHR data enables accurate early prediction of ED disposition, providing hospital decision-makers with early notice of inpatient demand to support proactive bed allocation, staffing coordination, and surge management.

Trial registration

Not applicable.