Construction And Comparison Of Risk Assessment Models For Rupture Of Anterior Choroidal Artery Aneurysms

PartⅠ:Nomograms for Assessing the Rupture Risk of Anterior Choroid Artery Aneurysms Relying upon Clinical,Morphological,and Hemodynamic FeaturesObjective:The objective is to develop nomograms for assessing the risk of rupture in unruptured aneurysms of the anterior choroidal artery（ACh A）.Methods:The clinical baseline,morphological,and haemodynamic characteristics of98 ACh A aneurysms were retrospectively collected between May 2017 and November2022,with 44 ruptured and 53 unruptured aneurysms,respectively.At the time of presentation to the hospitals,patients were divided into ruptured and unruptured groups based on the status of their ACh A aneurysm.To exclude multicollinearity interference,the variance inflation factor（VIF）was calculated,and parameters with VIF 4.0 were retained.Subsequently,the remaining parameters were put through to univariate logistic regression analysis,and those with P<0.20 were subjected to develop regression equations.Notably,the variables were evaluated with backward stepwise regression using the minimum Akaike information criterion.Our researchers established two prediction models.Finally,we evaluated the model’s discrimination,calibration,clinical benefit and models’performance,respectively.Results:Based on the minimum Akaike information criterion,four best candidates were finally obtained for the construction of the assessment model,namely smoking history,size ratio,normalized wall shear stress,and average oscillatory shear index.Model performance and comparison:1)Area under curve（AUC）of 0.795 and 0.706 for model 1and model 2,respectively;2)Calibration plots showing a high degree of agreement between predicted and measured values;3)Clinical decision curves suggesting that Model1,which includes haemodynamic parameters,has greater clinical benefit;4)Additionally,the net reclassification index,integrated discrimination improvement,and AUC indicated that model 1 with three dimensional parameters performs significantly better than model 2with two dimensional parameters..Conclusions:We developed two assessment models for the risk of rupture in ACh A aneurysms.Model 1 had greater accuracy,calibration,and clinical utility,while model 2had the advantage of time-saving.These nomogram models may offer valued tools for personalized risk assessment of unruptured ACh A aneurysms.PartⅡ:Rupture Risk Assessment for Anterior Choroid Artery Aneurysms Using Interpretable Machine Learning on Multidimensional DataObjective:Machine learning is currently being employed extensively in the risk assessment of intracranial aneurysm rupture,and it has produced promising outcomes.However,the vast majority of the models were built by including all intracranial aneurysms（IAs）,with only a very small percentage of anterior choroidal artery（ACh A）aneurysms.However,the construction of machine learning（ML）models with a limited number of parameters would lack interpretability,which may restrict the scalability of evaluation models in clinical contexts.Our objective is to construct interpretable machine learning models using multidimensional data for assessing the risk of ACh A aneurysm rupture.Methods:From December 2017 to February 2023,a total of 103 ACh A aneurysms were included in this study.Each patient included eight clinical baseline characteristics,seven aneurysm morphologies,and five hemodynamic parameters gathered.The model for risk assessment was established utilizing classical logistic regression（LR）and six machine learning techniques,namely extreme gradient boosting（XGBoost）,random forest（RF）,support vector machine（SVM）,K-nearest neighbor（KNN）,light gradient boosting machine（Light GBM）,multi-layer perception（MLP）.The models are compared in terms of discrimination,calibration,and clinical application,and Shapley additive explanations analysis is utilized to enhance the interpretability of the ideal machine learning models and disclose the“black box”underlying their predictions.Results:On univariate analysis,There were substantial gaps in comprehension between the two different groups in the 12 parameters of smoking,daughter sac,undulation index（UI）,nonspericity index（NSI）,ellipticity index（EI）,size ratio（SR）,aspect ratio（AR）,size,normalized wall shear stress（NWSS）,average oscillatory shear index(OSI_ave),low wall shear stress area（LSA）,relative residence time（all P<0.05）.Variables with significant univariate analysis results were submitted to a variance inflation factor（VIF）test,and ellipticity index（VIF>4）were removed from the model building factor cohort.In the inner validation cohort,the XGBoost model obtained the highest area under the receiver operating characteristic curve（0.839,95%CI:0.641～0.997）,higher than the Light GBM model（0.806,95%CI:0.598～0.973）,SVM model（0.798,95%CI:0.579～0.985）,KNN model（0.764,95%CI:0.547～0.967）,RF model（0.730,95%CI:0.486～0.963）,LR（0.713,95%CI:0.429～0.973）,MLP model（0.563,95%CI:0.287～0.829）.The XGBoost model was superior to the LR model by De Long’s test（P=0.024）.Shapley additive explanations analysis revealed that the XGBoost models with the best predictive ability were ranked in order of factor importance as NSI,SR,LSA,AR,size,smoking,OSI_ave,RRT,NWSS,UI,and daughter sac.Conclusions:This research highlights the potential of machine learning for assessing the rupture risk of ACh A aneurysms.Machine learning models led by the extreme gradient boosting method outperformed traditional logistic regression models overall.The Shapley additive explanations analysis improves the interpretability of machine learning models and highlights important risk factors associated with the risk of ACh A aneurysm rupture.These findings suggest that machine learning models can improve clinical decision-making in the assessment of ACh A aneurysm stability,potentially allowing patients to receive individualised and accurate treatment.