Scholarlink

Background <p>Escherichia coli bloodstream infection (<i>E. coli</i> BSI) is a global clinical challenge with markedly higher mortality once it progresses to septic shock. Current tools are impractical in primary care, creating an urgent clinical need for accessible, laboratory-based early prediction tools. Thus, we developed and validated an interpretable machine learning model using routine laboratory indicators acquired prior to septic shock to predict septic shock in patients with <i>E. coli</i> BSI.</p> Methods <p>A retrospective analysis was conducted on 719 patients with <i>E. coli</i> BSI treated at a tertiary referral center. The cohort was randomly divided into training (<i>n</i> = 503) and test (<i>n</i> = 216) sets. Feature selection was performed using univariate analysis, LASSO regression, and the Boruta algorithm. The chosen final model underwent 10-fold cross-validation, utilizing the area under the receiver operating characteristic curve (AUC) as the main metric for discrimination. To account for unequal class sizes, balanced accuracy was used. Bootstrap resampling with replacement was performed on all enrolled patients, alongside external validation, to assess the stability of the final model. Decision curve analysis (DCA) was performed to confirm the model’s clinical benefits. SHapley Additive exPlanations (SHAP) analysis was conducted to strengthen model interpretability.</p> Results <p>Five core features were selected as predictive factors: procalcitonin, total protein, chloride, platelet count, and total carbon dioxide. The Extreme Gradient Boosting model achieved the best predictive performance among all candidates, with an AUC of 0.807 in the test set, 0.836 in bootstrap validation, and a balanced accuracy of 0.741. The DCA indicated a substantial clinical net benefit. External validation confirmed the stable predictive efficacy of the model, with a consistent AUC of 0.807. SHAP analysis indicated that procalcitonin and total protein were strongly correlated with the onset of septic shock.</p> Conclusion <p>A robust and interpretable machine learning model constructed using routine laboratory indicators and validated through both bootstrap resampling and external validation was established to predict septic shock in patients with <i>E. coli</i> BSI. This approach provides a convenient and objective tool for risk assessment, enabling timely interventions and potentially improving outcomes in this high-mortality population.</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

An early prediction model of septic shock in patients with Escherichia coli bloodstream infection based on machine learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月27日

查看全文
  • Xinran Qiu,
  • Yi Wu,
  • Wenqi Ouyang,
  • Junlei Zhou,
  • Lele Zheng,
  • Guanhua Li,
  • Guohui Xue,
  • Jinshen Chu

摘要

Background

Escherichia coli bloodstream infection (E. coli BSI) is a global clinical challenge with markedly higher mortality once it progresses to septic shock. Current tools are impractical in primary care, creating an urgent clinical need for accessible, laboratory-based early prediction tools. Thus, we developed and validated an interpretable machine learning model using routine laboratory indicators acquired prior to septic shock to predict septic shock in patients with E. coli BSI.

Methods

A retrospective analysis was conducted on 719 patients with E. coli BSI treated at a tertiary referral center. The cohort was randomly divided into training (n = 503) and test (n = 216) sets. Feature selection was performed using univariate analysis, LASSO regression, and the Boruta algorithm. The chosen final model underwent 10-fold cross-validation, utilizing the area under the receiver operating characteristic curve (AUC) as the main metric for discrimination. To account for unequal class sizes, balanced accuracy was used. Bootstrap resampling with replacement was performed on all enrolled patients, alongside external validation, to assess the stability of the final model. Decision curve analysis (DCA) was performed to confirm the model’s clinical benefits. SHapley Additive exPlanations (SHAP) analysis was conducted to strengthen model interpretability.

Results

Five core features were selected as predictive factors: procalcitonin, total protein, chloride, platelet count, and total carbon dioxide. The Extreme Gradient Boosting model achieved the best predictive performance among all candidates, with an AUC of 0.807 in the test set, 0.836 in bootstrap validation, and a balanced accuracy of 0.741. The DCA indicated a substantial clinical net benefit. External validation confirmed the stable predictive efficacy of the model, with a consistent AUC of 0.807. SHAP analysis indicated that procalcitonin and total protein were strongly correlated with the onset of septic shock.

Conclusion

A robust and interpretable machine learning model constructed using routine laboratory indicators and validated through both bootstrap resampling and external validation was established to predict septic shock in patients with E. coli BSI. This approach provides a convenient and objective tool for risk assessment, enabling timely interventions and potentially improving outcomes in this high-mortality population.