文献速递神经外周血管CTDSA

基于CT血管造影血流动力学机器学习预测模型用于颅内动脉瘤破裂状态的开发与验证:一项中国多中心研究

Development and validation of machine learning prediction model based on computed tomography angiography-derived hemodynamics for rupture status of intracranial aneurysms: a Chinese multicenter study.
PubMed 文献 · Eur Radiol · 2020-04-29

该研究基于CT血管造影诊断并经侵袭性脑血管造影或夹闭手术确认的颅内动脉瘤患者,构建结合临床、动脉瘤形态和血流动力学参数的常规逻辑回归与机器学习模型预测未破裂颅内动脉瘤的破裂状态,并在内部和外部验证集中测试。训练集807例、内部验证集200例、外部验证集108例;常规逻辑回归模型1(临床)、2(临床+形态)、3(临床+形态+血流动力学)的AUC分别为0.608、0.765、0.886(均p<0.05);机器学习模型随机森林、多层感知机、支持向量机的AUC分别为0.871、0.851、0.863;常规逻辑回归、随机森林和支持向量机的AUC之间无差异(均p>0.05/6),而多层感知机的AUC低于常规逻辑回归(p=0.0055)。结论:血流动力学参数在模型预测性能中起重要作用;在整合临床、动脉瘤形态和血流动力学参数的破裂状态预测模型中,机器学习方法未能优于常规逻辑回归。

要点

  • 加入血流动力学参数可提高未破裂颅内动脉瘤破裂状态的预测性能。
  • 在整合临床、形态和血流动力学参数的模型中,机器学习算法未能优于常规逻辑回归。
  • 整合临床、动脉瘤形态和血流动力学参数的模型可能有助于选择最佳管理方案。

诊断标准与关键数值

  • 常规逻辑回归模型1(临床)AUC为0.608。
  • 常规逻辑回归模型2(临床和动脉瘤形态)AUC为0.765。
  • 常规逻辑回归模型3(临床、动脉瘤形态和血流动力学特征)AUC为0.886(均p<0.05)。
  • 机器学习模型随机森林、多层感知机、支持向量机的AUC分别为0.871、0.851、0.863。
  • 常规逻辑回归、随机森林和支持向量机的AUC之间无差异(均p>0.05/6),多层感知机的AUC低于常规逻辑回归(p=0.0055)。

临床应用与阅片要点

  • 在预测未破裂颅内动脉瘤破裂状态时,可考虑在临床和形态学参数基础上加入血流动力学参数以提高模型性能。
  • 机器学习算法在整合临床、形态和血流动力学参数的破裂状态预测模型中未显示优于常规逻辑回归。
  • 整合临床、动脉瘤形态和血流动力学参数的模型可能有助于选择最佳管理方案。

局限性

  • 原文未报告局限性。

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)。

英文原文(摘要)· 点击展开

OBJECTIVES: To build models based on conventional logistic regression (LR) and machine learning (ML) algorithms combining clinical, morphological, and hemodynamic information to predict individual rupture status of unruptured intracranial aneurysms (UIAs), afterwards tested in internal and external validation datasets.

METHODS: Patients with intracranial aneurysms diagnosed by computed tomography angiography and confirmed by invasive cerebral angiograph or clipping surgery were included. The prediction models were developed based on clinical, aneurysm morphological, and hemodynamic parameters by conventional LR and ML methods.

RESULTS: The training, internal validation, and external validation cohorts were composed of 807 patients, 200 patients, and 108 patients, respectively. The area under curves (AUCs) of conventional LR models 1 (clinical), 2 (clinical and aneurysm morphological), and 3 (clinical, aneurysm morphological and hemodynamic characteristics) were 0.608, 0.765, and 0.886, respectively (all p < 0.05). The AUCs of ML models using random forest (RF), multilayer perceptron (MLP), and support vector machine (SVM) were 0.871, 0.851, and 0.863, respectively. There were no difference among AUCs of conventional LR, RF, and SVM (all p > 0.05/6), while the AUC of MLP was lower than that of conventional LR (p = 0.0055).

CONCLUSION: Hemodynamic parameters play an important role in the prediction performance of the models. ML methods cannot outperform conventional LR in prediction models for rupture status of UIAs integrating clinical, aneurysm morphological, and hemodynamic parameters.

KEY POINTS: • The addition of hemodynamic parameters can improve prediction performance for rupture status of unruptured intracranial aneurysms. • Machine learning algorithms cannot outperform conventional logistic regression in prediction models for rupture status integrating clinical, aneurysm morphological, and hemodynamic parameters. • Models integrating clinical, aneurysm morphological, and hemodynamic parameters may help choose the optimal management.

来源:Eur Radiol,查看原文。版权归原作者所有。

引用

  1. Chen G, Lu M, Shi Z, et al. Development and validation of machine learning prediction model based on computed tomography angiography-derived hemodynamics for rupture status of intracranial aneurysms: a Chinese multicenter study. Eur Radiol. 2020. PMID: 32350658. doi:10.1007/s00330-020-06886-7

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。