Scholarlink

Objective <p>This study aimed to develop an internally validated machine-learning model for predicting chronic hydrocephalus after aneurysmal subarachnoid hemorrhage (aSAH) using early cerebrospinal fluid (CSF), peripheral inflammatory, and CT imaging markers.</p> Materials and methods <p>This single-center retrospective cohort study included 524 patients with aSAH. Chronic hydrocephalus was defined according to radiological and clinical criteria within 6&#xa0;months after ictus or index admission. The cohort was randomly split into training and validation cohorts at a 7:3 ratio. Candidate predictors were selected using LASSO regression and assessed for multicollinearity by variance inflation factors. Eight machine-learning models, including logistic regression, KNN, neural network, random forest, GBM, XGBoost, SVM, and AdaBoost, were developed using repeated tenfold cross-validation with five repeats. Model performance was evaluated in terms of discrimination with 95% confidence intervals, calibration, and clinical utility. Pairwise AUC comparisons were conducted using the DeLong test, and SHAP analysis was used for model interpretability. As a supplementary analysis, SMOTE was applied to the training data to examine the robustness of model performance under class-imbalance correction. An online risk calculator was developed to facilitate individualized risk estimation. Missing-data sensitivity analyses were performed by comparing included and excluded patients using standardized mean differences.</p> Results <p>LASSO identified key predictors, including modified Fisher grade, CSF WBC, CSF protein, CSF chloride, SIRI, PAR, and ventricular/callosal-angle imaging measures. Among the evaluated models, KNN demonstrated numerically favorable overall performance in internal validation, with an AUC of 0.821 (95% CI, 0.735–0.906) and a Brier score of 0.089 in the validation cohort. In the training cohort, KNN achieved an AUC of 0.946 (95% CI, 0.925–0.966) and a Brier score of 0.052. Pairwise DeLong tests showed no statistically significant differences in validation-cohort AUCs between models (all <i>P</i> &gt; 0.05). Decision curve analysis demonstrated net benefit across threshold probabilities of 0.2–0.8 in the validation cohort. An online calculator is available at <a href="https://ycyy-hydrocephalus-predictive-model.shinyapps.io/workrun10/">https://ycyy-hydrocephalus-predictive-model.shinyapps.io/workrun10/</a>.</p> Conclusions <p>The KNN model demonstrated preliminary potential for early prediction of chronic hydrocephalus after aSAH, with numerically favorable overall performance in internal validation. However, given the retrospective single-center design, potential overfitting, and lack of external validation, the model should be regarded as exploratory. External validation in independent multicenter cohorts and prospective validation are required before clinical implementation.</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

A multimodal machine-learning model incorporating peripheral blood inflammatory markers, cerebrospinal fluid profiles, and preoperative CT parameters to predict chronic hydrocephalus following aneurysmal subarachnoid hemorrhage

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Yunchong Xiao,
  • Weichong Zhou,
  • Hui Shi,
  • Xingfu Liao,
  • Xilu Yu,
  • anyi Liu,
  • Long Chen,
  • Hai Su

摘要

Objective

This study aimed to develop an internally validated machine-learning model for predicting chronic hydrocephalus after aneurysmal subarachnoid hemorrhage (aSAH) using early cerebrospinal fluid (CSF), peripheral inflammatory, and CT imaging markers.

Materials and methods

This single-center retrospective cohort study included 524 patients with aSAH. Chronic hydrocephalus was defined according to radiological and clinical criteria within 6 months after ictus or index admission. The cohort was randomly split into training and validation cohorts at a 7:3 ratio. Candidate predictors were selected using LASSO regression and assessed for multicollinearity by variance inflation factors. Eight machine-learning models, including logistic regression, KNN, neural network, random forest, GBM, XGBoost, SVM, and AdaBoost, were developed using repeated tenfold cross-validation with five repeats. Model performance was evaluated in terms of discrimination with 95% confidence intervals, calibration, and clinical utility. Pairwise AUC comparisons were conducted using the DeLong test, and SHAP analysis was used for model interpretability. As a supplementary analysis, SMOTE was applied to the training data to examine the robustness of model performance under class-imbalance correction. An online risk calculator was developed to facilitate individualized risk estimation. Missing-data sensitivity analyses were performed by comparing included and excluded patients using standardized mean differences.

Results

LASSO identified key predictors, including modified Fisher grade, CSF WBC, CSF protein, CSF chloride, SIRI, PAR, and ventricular/callosal-angle imaging measures. Among the evaluated models, KNN demonstrated numerically favorable overall performance in internal validation, with an AUC of 0.821 (95% CI, 0.735–0.906) and a Brier score of 0.089 in the validation cohort. In the training cohort, KNN achieved an AUC of 0.946 (95% CI, 0.925–0.966) and a Brier score of 0.052. Pairwise DeLong tests showed no statistically significant differences in validation-cohort AUCs between models (all P > 0.05). Decision curve analysis demonstrated net benefit across threshold probabilities of 0.2–0.8 in the validation cohort. An online calculator is available at https://ycyy-hydrocephalus-predictive-model.shinyapps.io/workrun10/.

Conclusions

The KNN model demonstrated preliminary potential for early prediction of chronic hydrocephalus after aSAH, with numerically favorable overall performance in internal validation. However, given the retrospective single-center design, potential overfitting, and lack of external validation, the model should be regarded as exploratory. External validation in independent multicenter cohorts and prospective validation are required before clinical implementation.