机器学习模型对脑室-腹腔分流术治疗颅脑创伤后脑积水患者预后的预测价值

Predictive value of machine learning model for the prognosis of patients with hydrocephalus after craniocerebral injury treated by ventriculoperitoneal shunt

目的探讨机器学习模型在预测脑室-腹腔分流术(VPS)治疗颅脑创伤后脑积水患者预后中的价值。方法回顾性分析2013年1月至2019年7月复旦大学附属华山医院神经外科收治的因创伤后脑积水行VPS患者的临床资料,共94例。比较分流术前与分流术后出院时最后一次格拉斯哥昏迷评分(GCS)变化,确定进步组(GCS提高≥1分)和未进步组(GCS提高<1分)。采用传统的统计学方法,比较两组患者的年龄、性别、住院时间,颅内感染、后颅窝病变、脑室出血、颅骨缺损、脑室压力及脑积水类型(低压性脑积水和正常压力性脑积水)的差异。采用Python 3.6和scikit-learn 0.23工具包进行机器学习分析,采用logistic回归算法作为机器学习分析的基本算法,随机进行训练测试拆分,其中70%的数据(65例)为训练集,30%的数据(29例)为测试集。分析上述危险因素对分流术后患者预后的预测作用。结果传统统计学结果显示,与未进步组比较,进步组患者的年龄大、分流术前的GCS低,差异均有统计学意义(均P<0.05);而颅内感染、后颅窝病变、脑室出血及颅骨缺损的发生率,以及住院时间、脑室压力、低压性脑积水,两组比较差异均无统计学意义(均P>0.05)。机器学习模型预测结果显示,在训练集预测VPS术后临床进步的准确率为(67.4±20.0)%,曲线下面积(AUC)为0.69±0.18;而测试集分别为62.1%、0.71。与临床诊断比较,基于机器学习模型预测VPS治疗脑创伤后脑积水不良预后的总体准确率为68.1%(64/94),灵敏度为71.4%(30/42)、特异度为65.4%(34/52),AUC为0.89(χ^(2)=12.600,P<0.001)。机器学习预测模型的权重分析显示,低压性脑积水占50.8%、分流术前的GCS占22.7%、颅内感染占13.2%、颅骨缺损占9.7%。结论与传统统计学方法比较,基于机器学习的预测模型可捕捉到统计效应更微小的临床指标,对颅脑创伤后脑积水不良预后的预测价值更大。

Objective To explore the value of machine learning model in predicting the prognosis of patients with hydrocephalus after craniocerebral injury treated by ventriculoperitoneal shunt(VPS).Methods A retrospective analysis was conducted on the clinical data of 94 patients who underwent VPS due to hydrocephalus after craniocerebral injury admitted to the Department of Neurosurgery of Huashan Hospital,Fudan University from January 2013 to July 2019.The last Glasgow coma score(GCS)at discharge after shunting was compared to preoperative GCS to determine the progressive group and the non-progressive group.Traditional statistical methods were used to compare the age,gender,length of hospital stay,intracranial infection,posterior fossa lesion,ventricular hemorrhage,skull defect,ventricular pressure and types of hydrocephalus(low pressure hydrocephalus and normal pressure hydrocephalus)between the two groups.Python 3.6 and scikit-learn 0.23 toolkit were used for machine learning analysis,logistic regression algorithm was used as the basic algorithm for machine learning analysis,and randomly split training and testing were performed.In this series,70%of the data(65 cases)were the training set,and the remaining 30%of the data(29 cases)were the test set.We then analyzed the predictive effect of the above risk factors on the prognosis of patients after shunt.Results Traditional statistical results showed that compared with the patients in non-progressive group,those in progressive group were older and had a lower GCS score before VPS,and the differences were statistically significant(both P<0.05);while there was no significant difference in the incidence of intracranial infection,posterior fossa lesion,intraventricular hemorrhage,skull defect,length of hospital stay,ventricular pressure,or presence of low-pressure hydrocephalus between the two groups(all P>0.05).The prediction results of the machine learning model showed that the accuracy rate of predicting clinical progress after VPS in the training set was 67.4±20.0%,and the AUC was 0.69±0.18,and those in the test sets were 62.1%and 0.71 respectively.Compared with clinical diagnosis,the overall accuracy rate of predicting the poor prognosis of hydrocephalus after brain injury based on machine learning risk factors was 68.1%(64/94),the sensitivity was 71.4%(30/42),the specificity was 65.4%(34/52),and AUC was 0.89(χ^(2)=12.600,P<0.001).The weight analysis of the machine learning prediction model showed that low-pressure hydrocephalus accounted for 50.8%,pre-shunt GCS score accounted for 22.7%,intracranial infection accounted for 13.2%,and skull defect accounted for 9.7%.Conclusions Compared with traditional statistical methods,the prediction model based on machine learning can capture clinical indicators with smaller statistical effects,and has greater predictive value for the poor prognosis of hydrocephalus after brain injury.