文献速递脑血管病外科神经重症

机器学习模型预测破裂颅内动脉瘤后血管痉挛的系统综述与荟萃分析

Machine learning models for predicting vasospasm following ruptured intracranial aneurysms: a systematic review and meta-analysis.
PubMed 文献 · Acta Neurochir (Wien) · 2025-12-03

深度学习灵敏度最高(均值97.6%),AUC-ROC达0.97,在灵敏度上优于回归、集成和SVM方法(p=0.003);但方法学异质性、外部验证有限及缺乏前瞻性试验阻碍临床 adoption。

要点

  • 深度学习预测血管痉挛灵敏度达97.6%,AUC-ROC为0.97。
  • SVM模型阴性预测值最高(85%),集成与回归方法阳性预测值更优。
  • 偏倚风险低至中等,主要关注患者选择与缺乏外部验证。

核心结论

  • 机器学习模型尤其是深度学习和集成方法在预测aSAH后血管痉挛方面展现出有希望的准确性。
  • 这些工具有望实现更早的个性化干预,但方法学异质性、外部验证有限及缺乏前瞻性试验目前阻碍临床采用。

研究设计与样本

  • 系统综述与荟萃分析,遵循PRISMA 2020指南,检索PubMed和Embase数据库。

主要发现

  • 深度学习灵敏度最高(均值97.6%),AUC-ROC为0.97,在灵敏度上优于回归、集成和SVM方法(p=0.003),但在特异度或AUC上无差异。
  • SVM模型阴性预测值最高(85%),集成和回归方法阳性预测值更优。
  • 深度学习在不同队列类型中均提供高准确性和泛化性,但阳性预测值变异性较大。
  • 双变量分析证实人工神经网络和随机森林模型实现了有利的灵敏度-特异度权衡。
  • 偏倚风险低至中等,多数关注点与患者选择及缺乏外部验证相关。

临床怎么用

  • ML模型尤其深度学习和集成方法可用于早期识别aSAH后血管痉挛高风险患者,辅助个性化干预决策。
  • 临床 adoption 前需考虑方法学异质性和外部验证缺乏的局限,不宜直接替代现有监测策略。

局限性

  • 方法学异质性较大。
  • 外部验证有限。
  • 缺乏前瞻性试验。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: Cerebral vasospasm remains a leading cause of delayed cerebral ischemia following aneurysmal subarachnoid hemorrhage (aSAH). Despite advances in critical care, current monitoring strategies are reactive and non-personalized. Machine learning (ML) has emerged as a promising tool to anticipate vasospasm risk. METHODS: A systematic review and meta-analysis were performed following PRISMA 2020 guidelines. PubMed and Embase databases were searched for studies applying ML algorithms to predict clinical or radiological vasospasm. Data were pooled using bivariate and proportional meta-analyses and their quality was assessed with the PROBAST tool. RESULTS: Twelve studies (2011-2025) encompassing 25 ML models were included. Deep learning achieved the highest sensitivity (mean: 97.6%) and AUC-ROC (0.97), outperforming regression, ensemble, and SVM methods in sensitivity (p = 0.003) but not in specificity or AUC. SVM models showed the highest NPV (85%), while ensemble and regression methods had superior PPV. Across cohort types, deep learning consistently delivered high accuracy and generalizability, although with greater PPV variability. Bivariate analysis confirmed that artificial neural networks and random forest models achieved favorable sensitivity-specificity trade-offs. Risk of bias was low to moderate, with most concerns related to patient selection and lack of external validation. CONCLUSION: ML models, particularly deep learning and ensemble methods, demonstrate promising accuracy in predicting vasospasm after aSAH. These tools may enable earlier, personalized interventions; however, methodological heterogeneity, limited external validation, and lack of prospective trials currently hinder clinical adoption.

来源:Acta Neurochir (Wien),查看原文。版权归原作者所有。

引用

  1. Palermo M, D'Arrigo S, Olivi A, et al. Machine learning models for predicting vasospasm following ruptured intracranial aneurysms: a systematic review and meta-analysis. Acta Neurochir (Wien). 2025. PMID: 41339534. doi:10.1007/s00701-025-06725-y

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。