Scholarlink

<p>Respiratory complications are major contributors to morbidity and mortality in trauma patients, yet conventional predictive models remain limited in scope and performance. We developed and validated an ensemble machine learning model that integrates both prehospital and in-hospital clinical variables to predict respiratory complications in trauma patients. We developed and internally validated machine learning models using data from the Korea Trauma Data Bank, comprising records from 19 major trauma centers in Korea between 2017 and 2022 (discovery; <i>n</i> = 48,376). For external validation, data from four additional trauma centers added in 2023 (external validation; <i>n</i> = 2,010) were used. Trauma patients were identified using S or T codes in accordance with the 7th Korean Standard Classification of Diseases. Respiratory complications were defined as a composite outcome including acute respiratory distress syndrome, pneumonia, and unplanned intubation. The models were trained using 19 pre-hospital and in-hospital variables, and the final prediction model was constructed by ensembling the top-performing models. Model interpretability was achieved through Shapley Additive Explanations (SHAP). Lastly, the predicted probabilities were categorized into tertiles (T1, T2, and T3), and their association with in-hospital mortality was assessed using logistic regression analysis. Among 48,376 trauma patients in the discovery cohort, the final soft-voting ensemble model combining adaptive boosting, gradient boosting machine, and logistic regression achieved an area under receiver operating characteristic curve of 0.834 in discovery cohort and 0.839 in external validation cohort. SHAP analysis identified pre-hospital pulse, Injury Severity Score, and age as the most influential predictors of respiratory complications. Higher predicted risk of respiratory complications was significantly associated with in-hospital mortality, with adjusted odds ratios rising across predicted risk tertiles (T1: 2.59 [95% CI, 2.09–3.22]; T2: 2.77 [2.21–3.47]; and T3: 3.91 [2.30–5.11]). The proposed ensemble model exhibited high accuracy and generalizability in predicting respiratory complications and mortality risk among trauma patients. By using both pre-hospital and in-hospital clinical data, the model offers a potentially valuable tool for early triage and intervention in trauma care. Prospective validation is needed to evaluate its clinical utility in real-world trauma settings.</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

Development and validation of a prediction model for respiratory complications and in-hospital mortality in trauma patients

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月9日

查看全文
  • Do Wan Kim,
  • Dongjin Yeo,
  • Juyeong Kim,
  • Yerin Hwang,
  • Hyunjee Kim,
  • Seung Ha Hwang,
  • Jaeyu Park,
  • Jinseok Lee,
  • Jaehyeong Cho,
  • Selin Woo,
  • Junepill Seok,
  • Byungchul Yu,
  • Youngmin Kim,
  • Sebeom Jeon,
  • Dong Keon Yon,
  • Wu Seong Kang

摘要

Respiratory complications are major contributors to morbidity and mortality in trauma patients, yet conventional predictive models remain limited in scope and performance. We developed and validated an ensemble machine learning model that integrates both prehospital and in-hospital clinical variables to predict respiratory complications in trauma patients. We developed and internally validated machine learning models using data from the Korea Trauma Data Bank, comprising records from 19 major trauma centers in Korea between 2017 and 2022 (discovery; n = 48,376). For external validation, data from four additional trauma centers added in 2023 (external validation; n = 2,010) were used. Trauma patients were identified using S or T codes in accordance with the 7th Korean Standard Classification of Diseases. Respiratory complications were defined as a composite outcome including acute respiratory distress syndrome, pneumonia, and unplanned intubation. The models were trained using 19 pre-hospital and in-hospital variables, and the final prediction model was constructed by ensembling the top-performing models. Model interpretability was achieved through Shapley Additive Explanations (SHAP). Lastly, the predicted probabilities were categorized into tertiles (T1, T2, and T3), and their association with in-hospital mortality was assessed using logistic regression analysis. Among 48,376 trauma patients in the discovery cohort, the final soft-voting ensemble model combining adaptive boosting, gradient boosting machine, and logistic regression achieved an area under receiver operating characteristic curve of 0.834 in discovery cohort and 0.839 in external validation cohort. SHAP analysis identified pre-hospital pulse, Injury Severity Score, and age as the most influential predictors of respiratory complications. Higher predicted risk of respiratory complications was significantly associated with in-hospital mortality, with adjusted odds ratios rising across predicted risk tertiles (T1: 2.59 [95% CI, 2.09–3.22]; T2: 2.77 [2.21–3.47]; and T3: 3.91 [2.30–5.11]). The proposed ensemble model exhibited high accuracy and generalizability in predicting respiratory complications and mortality risk among trauma patients. By using both pre-hospital and in-hospital clinical data, the model offers a potentially valuable tool for early triage and intervention in trauma care. Prospective validation is needed to evaluate its clinical utility in real-world trauma settings.