Scholarlink

Background <p>Aneurysmal subarachnoid hemorrhage (aSAH) is a life-threatening cerebrovascular emergency. We built and validated a machine learning model integrating clinical and inflammatory indicators for early risk prediction.</p> Methods <p>This multicenter retrospective cohort study included 1,120 aSAH patients admitted between January 2022 and December 2024 across four tertiary hospitals for model development and 326 independent patients from the Second Xiangya Hospital for quasi-external validation. Twenty-eight candidate predictors were evaluated, encompassing clinical grading scales and inflammation- and nutrition-related biomarkers. Continuous variables were discretized into quartile-based categories to enhance interpretability and mitigate outlier effects. Synthetic minority oversampling (SMOTE) addressed outcome imbalance. Feature selection used a two-step process combining the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression, with variance inflation factor (VIF) analysis confirming the absence of collinearity. Six supervised algorithms were trained with tenfold cross-validation: logistic regression, neural network, random forest (RF), support vector machine (SVM), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost). Model performance was evaluated by discrimination, calibration, and decision curve analysis, and interpretability was assessed with Shapley additive explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME).</p> Results <p>The GBM model achieved the best performance, with an AUC of 0.895 (95% CI: 0.856–0.934) in internal validation and 0.864 (95% CI: 0.822–0.906) in quasi-external validation. Nine predictors were retained: procalcitonin, C-reactive protein-to-lymphocyte ratio (CLR), WFNS grade, systemic immune-inflammation index (SII), prognostic nutritional index (PNI), neutrophil-to-albumin ratio (NAR), Glasgow Coma Scale (GCS), platelet-to-lymphocyte ratio (PLR), and modified Fisher grade. A web-based calculator was implemented for individualized risk prediction.</p> Conclusions <p>The GBM-based model enables early prediction of poor short-term outcomes in aSAH, supporting timely clinical decision-making. Prospective multicenter validation is warranted to confirm its generalizability across diverse populations.</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

Machine learning-based prediction of short-term outcomes in aneurysmal subarachnoid hemorrhage: a multicenter study integrating clinical and inflammatory indicators

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月29日

查看全文
  • Weichong Zhou,
  • Cong Peng,
  • Peize Li,
  • Yufei Li,
  • Xingfu Liao,
  • Zhuo Wang,
  • Mingfeng Wang,
  • Yunchong Xiao,
  • Hai Su,
  • Hui Shi

摘要

Background

Aneurysmal subarachnoid hemorrhage (aSAH) is a life-threatening cerebrovascular emergency. We built and validated a machine learning model integrating clinical and inflammatory indicators for early risk prediction.

Methods

This multicenter retrospective cohort study included 1,120 aSAH patients admitted between January 2022 and December 2024 across four tertiary hospitals for model development and 326 independent patients from the Second Xiangya Hospital for quasi-external validation. Twenty-eight candidate predictors were evaluated, encompassing clinical grading scales and inflammation- and nutrition-related biomarkers. Continuous variables were discretized into quartile-based categories to enhance interpretability and mitigate outlier effects. Synthetic minority oversampling (SMOTE) addressed outcome imbalance. Feature selection used a two-step process combining the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression, with variance inflation factor (VIF) analysis confirming the absence of collinearity. Six supervised algorithms were trained with tenfold cross-validation: logistic regression, neural network, random forest (RF), support vector machine (SVM), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost). Model performance was evaluated by discrimination, calibration, and decision curve analysis, and interpretability was assessed with Shapley additive explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME).

Results

The GBM model achieved the best performance, with an AUC of 0.895 (95% CI: 0.856–0.934) in internal validation and 0.864 (95% CI: 0.822–0.906) in quasi-external validation. Nine predictors were retained: procalcitonin, C-reactive protein-to-lymphocyte ratio (CLR), WFNS grade, systemic immune-inflammation index (SII), prognostic nutritional index (PNI), neutrophil-to-albumin ratio (NAR), Glasgow Coma Scale (GCS), platelet-to-lymphocyte ratio (PLR), and modified Fisher grade. A web-based calculator was implemented for individualized risk prediction.

Conclusions

The GBM-based model enables early prediction of poor short-term outcomes in aSAH, supporting timely clinical decision-making. Prospective multicenter validation is warranted to confirm its generalizability across diverse populations.