Development of a pre-discharge model for 1-year post-discharge all-cause mortality after endovascular treatment for aneurysmal subarachnoid haemorrhage using LASSO–Boruta feature selection

To develop a predischarge model for predicting 1-year post-discharge all-cause mortality in patients with aneurysmal subarachnoid haemorrhage (aSAH) following endovascular treatment (EVT). We retrospectively analysed 947 patients with aSAH who were discharged alive between April 2021 and April 2023 from four neurointerventional centres in China as the training cohort. Candidate variables were selected using the least absolute shrinkage and selection operator (LASSO) combined with the Boruta algorithm. Based on these features, six models – logistic regression (LR), XGBoost, random forest (RF), AdaBoost, decision tree, and gradient boosting decision tree (GBDT) – were developed and compared. The optimal model was selected by the area under the receiver operating characteristic curve (AUC). The external validation cohort comprised 692 aSAH patients discharged alive between April 2023 and April 2024 from two additional centres. Model performance was evaluated using AUC, calibration curves, and decision curve analysis (DCA). Given the imbalanced outcome distribution, we applied the Synthetic Minority Over-sampling Technique (SMOTE) to further assess model generalisability. Among 1,639 patients alive at discharge, 67 (4.1%) died within 1 year. LASSO and Boruta jointly identified five key predictors for model construction: age, modified World Federation of Neurosurgical Societies (mWFNS) grade, ICU length of stay (ICU-LOS), C-reactive protein (CRP), and monocyte-to-HDL ratio (MHR). The random forest achieved the best discrimination in training set and remained strong in external validation cohorts.Moreover, SMOTE training yielded further improvements in generalisability. Random forest model enables individualised pre-discharge risk stratification and may guide perioperative management.

本研究旨在构建一款出院前预测模型，用于预测血管内治疗（endovascular treatment, EVT）后动脉瘤性蛛网膜下腔出血（aneurysmal subarachnoid haemorrhage, aSAH）患者出院后1年内的全因死亡风险。本研究回顾性纳入2021年4月至2023年4月期间，来自中国4家神经介入中心的947例存活出院的aSAH患者作为训练队列。采用最小绝对收缩和选择算子（least absolute shrinkage and selection operator, LASSO）联合Boruta算法筛选候选变量。基于筛选得到的特征，本研究构建并比较了6种模型：逻辑回归（logistic regression, LR）、XGBoost、随机森林（random forest, RF）、AdaBoost、决策树以及梯度提升决策树（gradient boosting decision tree, GBDT）。以受试者工作特征曲线下面积（area under the receiver operating characteristic curve, AUC）作为指标筛选最优模型。外部验证队列纳入2023年4月至2024年4月期间，来自另外2家中心的692例存活出院的aSAH患者。采用AUC、校准曲线以及决策曲线分析（decision curve analysis, DCA）评估模型性能。鉴于结局分布存在不平衡性，本研究采用合成少数类过采样技术（Synthetic Minority Over-sampling Technique, SMOTE）进一步评估模型的泛化能力。在1639例存活出院的患者中，共有67例（4.1%）在出院后1年内死亡。LASSO与Boruta算法联合筛选出5个构建模型的关键预测因子：年龄、改良世界神经外科医师联盟（modified World Federation of Neurosurgical Societies, mWFNS）分级、ICU住院时长（ICU-LOS）、C反应蛋白（C-reactive protein, CRP）以及单核细胞-高密度脂蛋白比值（monocyte-to-HDL ratio, MHR）。随机森林在训练队列中展现出最优的区分能力，且在外部验证队列中仍保持良好的性能。此外，采用SMOTE进行训练可进一步提升模型的泛化能力。本研究构建的随机森林模型可实现个体化出院前风险分层，有望为围手术期管理提供指导。