Johnson Mark D, Elavarthi Pradyumna, Street Seth, Hoz Samer S, Ralescu Anca L, Prestigiacomo Charles J
Department of Neurosurgery, University of Cincinnati, Cincinnati, Ohio, United States.
Department of Computer Science, University of Cincinnati, Cincinnati, Ohio, United States.
Surg Neurol Int. 2025 Jul 18;16:298. doi: 10.25259/SNI_498_2025. eCollection 2025.
With excitement in the medical community around artificial intelligence, machine learning (ML) techniques have been applied to correlate clinical and radiographic variables with intracranial aneurysm (IA) rupture status. In this study, we applied various ML techniques, including random forest (RF), XGBoost (XGB), support vector machines (SVM), and multi-layer perceptron (MLP), to predict IA rupture status.
The dataset consisted of 178 IAs each with 53 clinical and radiographic features for analysis. We removed features with high correlation (>0.8) with respect to the target variable to reduce redundancy. We applied grid search to fine-tune the hyperparameters for each model. Each model was evaluated across five iterations of 5-fold cross-validation. Overall performance metrics (accuracy, precision, recall, and F1-score) were extracted. The Wilcoxon signed-rank test was used to compare the area under the curve (AUC) scores between models.
The most common locations were internal carotid artery (42), anterior communicating artery (41), middle cerebral artery (32), and posterior communicating artery (25). The AUC for the RF (0.85) and XGB (0.76) models were significantly higher than those for the SVM (0.69) and MLP (0.65) models ( < 0.05). There was no statistical difference in accuracy between RF and XBG models ( = 0.144). Fractal dimension ranked as the most important feature for model performance across all models. Three-dimensional (3D) shape features made up 8 of the 15 most important features driving model performance.
Among the models, RF achieved the highest accuracy (85%) with balanced precision and recall. Across models 3D geometric features drove model performance, highlighting the importance of these features in predicting rupture status.
随着医学界对人工智能的热情高涨，机器学习（ML）技术已被应用于将临床和影像学变量与颅内动脉瘤（IA）破裂状态相关联。在本研究中，我们应用了各种ML技术，包括随机森林（RF）、极端梯度提升（XGB）、支持向量机（SVM）和多层感知器（MLP）来预测IA破裂状态。
数据集由178个IA组成，每个IA具有53个临床和影像学特征用于分析。我们去除了与目标变量高度相关（>0.8）的特征以减少冗余。我们应用网格搜索来微调每个模型的超参数。每个模型在5折交叉验证的五次迭代中进行评估。提取总体性能指标（准确率、精确率、召回率和F1分数）。使用Wilcoxon符号秩检验来比较模型之间的曲线下面积（AUC）分数。
最常见的部位是颈内动脉（42个）、前交通动脉（41个）、大脑中动脉（32个）和后交通动脉（25个）。RF（0.85）和XGB（0.76）模型的AUC显著高于SVM（0.69）和MLP（0.65）模型（<0.05）。RF和XBG模型之间的准确率没有统计学差异（=0.144）。分形维数在所有模型中排名为对模型性能最重要的特征。三维（3D）形状特征占驱动模型性能的15个最重要特征中的8个。
在这些模型中，RF在精确率和召回率平衡的情况下实现了最高准确率（85%）。在所有模型中，3D几何特征驱动模型性能，突出了这些特征在预测破裂状态中的重要性。
