9种机器学习模型预测幕上深部自发性脑出血早期血肿扩张及预后不良的比较

Comparative study of nine machine learning models for predicting early hematoma expansion and poor prognosis of deep supratentorial spontaneous intracerebral hemorrhage

目的比较9种机器学习模型对幕上深部自发性脑出血(SICH)患者发生早期血肿扩张及预后不良情况的预测性能。方法回顾性研究。纳入2015年1月—2019年5月4家医院幕上深部SICH患者420例。其中男275例、女145例,年龄25~90(61.0±12.9)岁。420例患者按照7∶3的比例,采用完全随机法分为训练集294例和验证集126例。患者在72 h内复查CT,若血肿体积比初始体积增长>6 mL或>33%,判定存在早期血肿扩张。采用改良的Rankin评分量表(mRS)评估预后,以mRS>3分判定为预后不良。比较训练集和验证集的基线资料。采用随机森林、极限梯度提升算法(XGboost)、梯度爬升决策树、自适应提升算法、朴素贝叶斯、logistic回归、支持向量机、K近邻、多层感知机9种机器学习算法对早期血肿扩张及预后不良分别构建预测模型;在训练集中,依据各模型的灵敏度和特异度绘制受试者操作特征曲线,采用3折交叉验证取曲线下面积(AUC),比较各模型对早期血肿扩张及预后不良情况的预测性能,并在验证集测试模型的可靠性。结果训练集和验证集患者基线资料比较差异均无统计学意义(P值均>0.05)。420例患者中,早期脑血肿扩张的患者有117例(27.86%);399例患者获随访,其中预后不良的患者有210例(52.63%)。预测早期血肿扩张:训练集中,9种机器学习模型的AUC为0.590~0.685,其中以XGboost模型最高,AUC为0.685±0.024;在验证集中,XGboost模型AUC为0.686[95%可信区间(CI)0.578~0.721]。预测预后不良:9种机器学习模型的AUC为0.703~0.852,其中logistic回归模型最高,AUC为0.852±0.041;而在验证集中,logistic回归模型AUC为0.894(95%CI 0.862~0.912)。结论9种机器学习算法模型中,XGboost对幕上深部SICH早期血肿扩张的预测性能最佳,而logistic回归模型对预后不良的预测性能最高;对于不同临床结局的预测,应选用合适的机器学习模型。

Objective This study aimed to compare the predictive performance of nine machine learning models for early hematoma expansion(HE)and poor outcomes in patients with supratentorial deep intracerebral hemorrhage(SICH).Methods In this study,a retrospective study design was used.A total of 420 patients with SICH in four hospitals from January 2015 to May 2019 were included,275 male and 145 female with a mean age of(61.0±12.9)years.The included patients were divided into 294 in the training set and 126 in the validation set in a 7∶3 ratio using randomization.In reviewed CT within 72 h,the hematoma volume increased by>6 mL or>33%,which was identified as HE.The prognosis was evaluated using the modified Rankin scale(mRS),and a score of mRS>3 was considered as a poor outcome.The baseline characteristics of the training and test sets were compared,and then nine machine learning algorithms including random forest,extreme gradient boosting(XGboost),gradient boosting decision tree,adaptive boosting,Naive Bayes,logistic regression,support vector machines,K-nearest neighbor,multilayer perceptron were used to construct prediction models for early HE and poor outcomes.In the training set,subject operating characteristic curves(ROC)were plotted on the basis of the sensitivity and specificity of each model,and the area under the curve(AUC)of threefold cross-validation was used to compare the predictive performance,which was investigated in the test set.Results The differences in baseline characteristics between the training and test sets were not statistically significant(all P values>0.05).Among the 420 patients,early HE was observed in 117 patients(27.86%).Follow-up results were obtained in 399 patients,and 210 patients(52.63%)had poor outcomes.In the training set,under threefold cross-validation,AUCs of nine machine learning models for predicting early HE ranged from 0.590 to 0.685,and the XGboost model was the highest at 0.685±0.024.In the validation set,the AUC was 0.686,and the 95%confidence interval(CI)was 0.578-0.721.In predicting poor outcomes,AUCs of the nine machine learning models ranged from 0.703 to 0.852,and the logistic regression model was the highest at 0.852±0.041.In the validation set,the AUC of the logistic regression model was 0.894(95%CI 0.862-0.912).Conclusion Among the nine machine learning models,XGboost has the best predictive performance for early HE of deep supratentorial SICH,whereas the logistic regression model has the highest predictive performance for poor outcomes.Furthermore,in predicting different clinical outcomes,an appropriate machine learning model should be selected.