基于临床与CT影像特征的可解释性机器学习模型预测脑出血早期血肿扩大的价值  

The Value of Interpretable Machine Learning Model Based on Clinical and CT Image Characteristics to Predict Early Hematoma Enlargementin Intracerebral Hemorrhage

作　　者：过永 张晓金 张林杰 徐家军 朱先锋 董明松 张虎 GUO Yong;ZHANG Xiaojin;ZHANG Linjie(Department of Radiology,Wuhu Second People's Hospital,Wuhu,Anhui Province 241000,P.R.China)

机构地区：[1]芜湖市第二人民医院放射科,241000

出　　处：《临床放射学杂志》2025年第5期814-820,共7页Journal of Clinical Radiology

基　　金：安徽省芜湖市皖南医学院校科研项目(编号:WK2023JXYY131)。

摘　　要：目的探讨基于临床与CT影像特征的可解释性多种机器学习(ML)算法、并构建选择最优输出模型预测原发性脑出血(PICH)早期血肿进展的价值。方法回顾性分析芜湖市第二人民医院260例PICH患者的临床、实验室检查、CT影像学资料,根据初始血肿体积和复查血肿体积的差值,分为进展组和稳定组。将所有患者按6∶4随机分为训练组(n=156)与验证组(n=104)。通过单因素与多因素Logistic回归分析筛选预测PICH早期血肿进展的独立危险因素。利用逻辑回归(LR)、K最近邻(KNN)、决策树(DT)、朴素贝叶斯(NB)、支持向量机(SVM)、随机森林(RF)6种不同ML算法分别构建预测模型,使用受试者工作特征曲线(ROC)曲线下面积(AUC)评估其性能。通过比较不同ML算法的性能,选出最佳模型,并利用Shapley加法解释方法(SHAP)值对该模型进行可视化,解释各个特征对预测结果的贡献。结果单因素与多因素Logistic回归分析显示形态评分、混合征、低密度征和糖尿病是预测PICH早期血肿扩大的独立危险因素。训练组ML模型的AUC为0.760~0.892;验证组ML模型的AUC为0.732~0.843。其中,RF模型在验证组中显示出最高的预测性能,其AUC值达到0.843,为最优输出模型,SHAP显示混合征是PICH早期血肿进展最重要的危险因素。结论结合临床与CT影像学特征的可解释性ML模型,可为PICH患者早期血肿进展提供更精确的评估。特别是RF模型在验证组中AUC值最大,显示出其在提升临床决策中具有较高价值。Objective To investigate multiple interpretable machine learning algorithms based on clinical and CT image characteristics and to construct and select the optimal output model to predict the early hematoma progression of primary cerebral hemorrhage.Methods The clinical,laboratory,and CT imaging data of 260 patients with primary cerebral hemorrhage from the Second People's Hospital of Wuhu City were analyzed retrospectively.Patients were divided into progressive and stable groups based on the difference between the initial and follow-up hematoma volumes.All patients were randomly divided into a training group(n=156)and a validation group(n=104)at a 6∶4 ratio.Univariate and multivariate logistic regression analyses were used to identify independent risk factors for predicting early hematoma progression in primary cerebral hemorrhage.Six different machine learning(ML)algorithms-Logistic regression(LR),K-nearest neighbor(KNN),decision tree(DT),naive Bayes(NB),support vector machine(SVM),and random forest(RF)-were employed to construct prediction models.The area under the receiver operating characteristic(ROC)curve(AUC)was used to evaluate model performance.The best model was selected by comparing the performance of different ML algorithms,and the Shapley Additive Explanations(SHAP)values were used to visualize the model and explain the contribution of each feature to the prediction results.Results Univariate and multivariate logistic regression analyses showed that morphological score,blend sign,low-density sign,and diabetes mellitus were independent risk factors for predicting early hematoma enlargement in primary cerebral hemorrhage.The AUC values of the ML models in the training group ranged from 0.760 to 0.892,while those in the validation group ranged from 0.732 to 0.843.The RF model demonstrated the highest predictive performance in the validation group,with an AUC value of 0.843.This model was selected as the optimal output model in this study.SHAP analysis revealed that the blend sign was the most important

关 键 词：体层摄影术 X线计算机 机器学习 原发性脑出血 血肿进展 SHAP 

分 类 号：R743.3[医药卫生—神经病学与精神病学]