基于深度神经网络原理构建宫腔镜下瘢痕部位妊娠物清除术时大出血风险预测模型

Based on the deep neural network principle,construct prediction model for massive bleeding risk during hysteroscopic removal of scar gestation

目的 探索基于深度神经网络原理构建宫腔镜下剖宫产瘢痕部位妊娠物清除术时大出血发生风险的预测模型,为预测大出血发生风险提供参考。方法 选取南宁市第一人民医院、南宁市第二人民医院2016年1月至2019年7月收治的200例剖宫产瘢痕妊娠(cesarean scar pregnancy, CSP)患者为数据集,收集所有入选患者的临床资料,随机选取数据集的60%为训练集,40%为测试集,对训练集和测试集中的出血组与对照组进行参数比较;根据模糊数学理论对数据集中的临床指标进行量化处理,采用R语言neuralnet包构建深度神经网络训练平台,预测行宫腔镜下瘢痕部位妊娠物清除术时大出血发生风险,并验证模型正确率和准确度。结果 训练集中出血组和对照组患者在住院次数、年龄、停经时间、术前凝血酶原时间(prothrombin time, PT)、活化部分凝血活酶时间(activated partial thromboplastin time, APTT)、凝血酶时间(thrombin time, TT)、纤维蛋白原(fibrinogen, Fib)、术前β-人绒毛膜促性腺激素(beta human chorionic gonadotropin, β-hCG)水平、B超孕囊最大径线、B超子宫瘢痕处肌层厚度、B超临床分型、手术时间的比较上,差异均有统计学意义(P<0.05);测试集中出血组和对照组患者在住院次数、术前PT、术前APTT、术前TT、术前Fib、B超孕囊最大径线、B超子宫瘢痕肌层厚度、B超临床分型及术前β-hCG水平的比较上,差异均有统计学意义(P<0.05)。利用深度神经网络构建的宫腔镜下瘢痕妊娠物清除术时大出血发生风险的预测模型中训练样集本120例,测试集样本80例,深度神经网络模型对数据分类的准确率均达到100.00%。在保证模型的稳定性和泛化性的前提下,通过模型混淆矩阵分析得出:训练集的准确度为0.925,敏感度为0.918,特异度为0.932,召回率为0.918,精确率为0.933。为保证模型预测的一致性和准确性,采用测试集进行验证,最终确定模型的准确度为0.915,敏感度为0.895,特异度为0.928,召回率为0.895,精确率为0.919。结论 本研究构建的深度神经网络模型对宫腔镜下瘢痕部位妊娠物清除术时大出血发生风险的预测准确率很高,分类性能优良,验证了B超子宫瘢痕肌层厚度≤0.2 cm、B超临床分型Ⅱ~Ⅲ型、剖宫产次数≥2次、术前Fib<3 g/L及停经时间>60 d是影响CSP患者行宫腔镜下瘢痕部位妊娠物清除术时大出血发生风险的危险因素。

ObjectiveTo explore the construction of a prediction model based on the principle of deep neural network to treat the risk of massive bleeding during cesarean scar pregnancy by hysteroscopic removal of pregnancy material,in order to predict cesarean scar pregnancy,to provide reference for the risk of massive hemorrhage.Methods 200 cases of cesarean scar pregnancy(CSP)patients predicting admitted to Nanning First People's Hospital and Nanning Second Peoples Hospital from January 2016 to July 2019 were selected as the data set,clinical data of all patients were collected,60%of the data set was randomly selected as the training set and 40%as the test set,the parameters of the bleeding group and the control group in the training set and sest set were compared.The clinical indicators in the data set were quantized according to the fuzzy mathematical theory,and the R language neuralnet package was used to build a deep neural network training plaform to predict the risk of massive bleeding during hysteroscopic removal of scar gestation in the treatment of CSP,and to verify the correctness and accuracy of the model.Results There were statistically significant differences in thrombin time,age,duration of menopause,preoperative prothrombin time(PT),and activated partial thromboplastin time(APTT),thrombin time(TT),fibrinogen(Fib),and preoperative beta human chorionic gonadotropin(β-hCG)level,maximum diameter of gestational sac by B-ultrasound,muscular layer thickness at uterine scar,clinical classification of B-ultrasound and operation time between the bleeding group and control group in the training set(P<0.05);There were statistically significant differences in hospitalization times,preoperative PT,preoperative APTT,preoperative TT,preoperative Fib,maximum diameter of gestational sac by Bultrasound,thickness of uterine cicatoid muscle layer by B-ultrasound,clinical classification of B-ultrasound and preoperativeβ-hCG level between the bleeding group and control group(P<0.05).In the prediction model of the risk of massive bleeding during hysteroscopic removal of scar gestation constructed by deep neural network,120 trained samples and 80 tested samples were included.The accuracy rate of data classification of deep neural network model reached 100.00%.On the premise of ensuring the stability and generalization of the model,the accuracy of the training set is 0.925,the sensitivity is 0.918,the specificity is 0.932,the recall rate is 0.918,and the accuracy rate is 0.933,through the model confusion matrix analysis.In order to ensure the consistency and accuracy of the model prediction,the test set was used to verify the accuracy of the model was 0.915,the sensitivity was 0.895,the specificity was 0.928,the recall rate was 0.895,and the accuracy rate was 0.919.Conclusions The deep neural network model successfully constructed in this study has a high prediction accuracy and excellent classification performance for the risk of massive bleeding during hysteroscopic removal of scar gestation.It was verified that B-ultrasound uterine cicatoid muscle thickness≤O.2 cm,B-ultrasound clinical classification type Ⅱ~Ⅲ,number of caesarean section≥2 times,preoperative Fib<3 g/L and duration of menstrual stoppage>60 days were risk factors affecting the risk of massive bleeding in CSP patients undergoing hysteroscopic removal of cicatoid gestation.