Scholarlink

<p>Intracerebral hemorrhage (ICH) carries high early mortality. To enable personalized decision-making, we developed and validated interpretable machine learning models for 30-day mortality prediction. This retrospective cohort study included patients with ICH extracted from the Medical Information Mart for Intensive Care (MIMIC) clinical database. Model development was performed using the MIMIC-IV (v3.1), while temporal validation was conducted on a distinct, non-overlapping cohort from the MIMIC-III CareVue subset (v1.4). Multiple imaging-free machine learning models were developed using routinely available clinical variables from the first 24&#xa0;h. Following data preprocessing and feature selection, the model development phase incorporated class weight balancing to address data imbalance and utilized Bayesian hyperparameter optimization. Discrimination, calibration, and decision curve analysis were assessed for all models. The final model was selected based on its superior performance and elucidated using SHapley Additive exPlanations (SHAP) to ensure transparency. Robustness and statistical reliability were evaluated through sensitivity analyses and sample size justification. Nine machine learning models were developed and evaluated in the MIMIC-IV cohort (<i>n</i> = 1,478; 1,034 for training, 444 for internal test). LightGBM was identified as the optimal model, demonstrating good discrimination (AUC: 0.859), adequate calibration (slope: 1.017; Brier score: 0.132), and potential clinical utility in decision curve analysis. Furthermore, on the temporal validation set (<i>n</i> = 339), the model maintained robust performance with an AUC of 0.811 and acceptable calibration (slope: 0.873; Brier score: 0.174). SHAP analysis enhanced clinical interpretability, and the model has been deployed as an open-access web tool for the early, individualized prediction. In this study, an interpretable LightGBM model demonstrated strong performance for 30-day mortality prediction in ICH, offering potential for individualized risk assessment and clinical integration. However, the lack of multicenter geographical external validation limits its generalizability, warranting further studies.</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 interpretable machine learning models to predict 30-day mortality in patients with intracerebral hemorrhage

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月12日

查看全文
  • Kai Yang,
  • Jianzhong Fang,
  • Xiufeng Zhang,
  • Yinghong Bai,
  • Hongming Ji

摘要

Intracerebral hemorrhage (ICH) carries high early mortality. To enable personalized decision-making, we developed and validated interpretable machine learning models for 30-day mortality prediction. This retrospective cohort study included patients with ICH extracted from the Medical Information Mart for Intensive Care (MIMIC) clinical database. Model development was performed using the MIMIC-IV (v3.1), while temporal validation was conducted on a distinct, non-overlapping cohort from the MIMIC-III CareVue subset (v1.4). Multiple imaging-free machine learning models were developed using routinely available clinical variables from the first 24 h. Following data preprocessing and feature selection, the model development phase incorporated class weight balancing to address data imbalance and utilized Bayesian hyperparameter optimization. Discrimination, calibration, and decision curve analysis were assessed for all models. The final model was selected based on its superior performance and elucidated using SHapley Additive exPlanations (SHAP) to ensure transparency. Robustness and statistical reliability were evaluated through sensitivity analyses and sample size justification. Nine machine learning models were developed and evaluated in the MIMIC-IV cohort (n = 1,478; 1,034 for training, 444 for internal test). LightGBM was identified as the optimal model, demonstrating good discrimination (AUC: 0.859), adequate calibration (slope: 1.017; Brier score: 0.132), and potential clinical utility in decision curve analysis. Furthermore, on the temporal validation set (n = 339), the model maintained robust performance with an AUC of 0.811 and acceptable calibration (slope: 0.873; Brier score: 0.174). SHAP analysis enhanced clinical interpretability, and the model has been deployed as an open-access web tool for the early, individualized prediction. In this study, an interpretable LightGBM model demonstrated strong performance for 30-day mortality prediction in ICH, offering potential for individualized risk assessment and clinical integration. However, the lack of multicenter geographical external validation limits its generalizability, warranting further studies.