Scholarlink

Background <p>Multiple organ dysfunction syndrome (MODS) is a severe complication of sepsis. This study aimed to develop and externally validate machine learning models based on routinely available clinical variables to predict the risk of new-onset MODS, occurring &gt; 24&#xa0;h after intensive care unit (ICU) admission, in patients with sepsis.</p> Methods <p>The models were developed using the Medical Information Mart for Intensive Care IV database and externally validated using the eICU Collaborative Research Database. Adult patients with sepsis admitted to an ICU for the first time were included, whereas those with MODS at admission or within the first 24&#xa0;h after ICU admission were excluded. The primary outcome was new-onset MODS occurring &gt; 24&#xa0;h after ICU admission. After feature selection using the least absolute shrinkage and selection operator regression, nine machine learning models were developed. Model performance was evaluated for discrimination, calibration, decision curve analysis, incremental predictive value, sensitivity analyses and SHapley Additive exPlanations (SHAP) interpretation.</p> Results <p>The final MIMIC-IV development cohort included 16,427 patients with sepsis, with 2381 patients developing new-onset MODS (14.5%). The external validation cohort comprised 1,678 patients, of whom 527 developed MODS (31.4%). Across the nine models, the area under the receiver operating characteristic curve (AUC) was 0.657–0.757 in the test set and 0.593–0.723 in the external validation cohort. The main ensemble learning models showed broadly comparable performance, with external validation AUCs ranging from 0.720 to 0.723. LightGBM achieved AUCs of 0.757 in the test set and 0.721 in the external validation cohort, with relatively stable discrimination and calibration. After removing the Sequential Organ Failure Assessment (SOFA) score, incremental value analysis and sensitivity analysis suggested that the machine learning models provided additional predictive information. SHAP analysis indicated that respiratory rate, SOFA score, mechanical ventilation, platelet count, blood urea nitrogen, vasoactive drug use and Glasgow Coma Scale score made substantial contributions to model prediction.</p> Conclusions <p>Machine learning models using routinely available clinical variables showed moderate predictive ability for new-onset MODS in patients with sepsis, and the main ensemble learning models demonstrated broadly comparable performance in the external validation cohort. Given the retrospective design and lack of prospective evaluation, these findings should be regarded as exploratory evidence for early risk stratification rather than support for immediate clinical implementation. Further prospective, multicentre studies are required to evaluate clinical utility.</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 external validation of a machine learning-based model for the early prediction of multiple organ dysfunction syndrome in critically ill patients with sepsis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月13日

查看全文
  • Jinbin Yang,
  • Linying Cai,
  • Xuyang Liu,
  • Kaihuan Zhou,
  • Junyu Lu,
  • Yegui Yang

摘要

Background

Multiple organ dysfunction syndrome (MODS) is a severe complication of sepsis. This study aimed to develop and externally validate machine learning models based on routinely available clinical variables to predict the risk of new-onset MODS, occurring > 24 h after intensive care unit (ICU) admission, in patients with sepsis.

Methods

The models were developed using the Medical Information Mart for Intensive Care IV database and externally validated using the eICU Collaborative Research Database. Adult patients with sepsis admitted to an ICU for the first time were included, whereas those with MODS at admission or within the first 24 h after ICU admission were excluded. The primary outcome was new-onset MODS occurring > 24 h after ICU admission. After feature selection using the least absolute shrinkage and selection operator regression, nine machine learning models were developed. Model performance was evaluated for discrimination, calibration, decision curve analysis, incremental predictive value, sensitivity analyses and SHapley Additive exPlanations (SHAP) interpretation.

Results

The final MIMIC-IV development cohort included 16,427 patients with sepsis, with 2381 patients developing new-onset MODS (14.5%). The external validation cohort comprised 1,678 patients, of whom 527 developed MODS (31.4%). Across the nine models, the area under the receiver operating characteristic curve (AUC) was 0.657–0.757 in the test set and 0.593–0.723 in the external validation cohort. The main ensemble learning models showed broadly comparable performance, with external validation AUCs ranging from 0.720 to 0.723. LightGBM achieved AUCs of 0.757 in the test set and 0.721 in the external validation cohort, with relatively stable discrimination and calibration. After removing the Sequential Organ Failure Assessment (SOFA) score, incremental value analysis and sensitivity analysis suggested that the machine learning models provided additional predictive information. SHAP analysis indicated that respiratory rate, SOFA score, mechanical ventilation, platelet count, blood urea nitrogen, vasoactive drug use and Glasgow Coma Scale score made substantial contributions to model prediction.

Conclusions

Machine learning models using routinely available clinical variables showed moderate predictive ability for new-onset MODS in patients with sepsis, and the main ensemble learning models demonstrated broadly comparable performance in the external validation cohort. Given the retrospective design and lack of prospective evaluation, these findings should be regarded as exploratory evidence for early risk stratification rather than support for immediate clinical implementation. Further prospective, multicentre studies are required to evaluate clinical utility.