基于机器学习和术前资料构建肝癌切除术中出血预测模型

Construction of an Intraoperative Bleeding Prediction Model for Hepatic Cancer Resection Based on Machine Learning and Preoperative Data

目的探索借助机器学习算法利用术前资料构建肝癌切除术中出血预测模型的可行性和应用价值。方法回顾性分析336例原发性肝癌手术患者的临床资料,按8∶2比例随机筛选出训练集268例和测试集68例。模型A采用基于实际手术时间的随机森林(RF)模型建立,模型B采用基于混合密度网络(MDN)模型的预测手术时间联合RF模型建立。二分类结局的单因素回归分析采用logistic回归模型。连续型结局变量筛选采用Lasso回归模型,模型预测精度采用均方根误差(RMSE)、平均绝对百分比误差(MAPE)等衡量;分类变量预测模型预测性能采用准确率、校准曲线、Brier评分、工作曲线下面积(AUC)等衡量;模型性能稳定性采用全部样本的bootstrap法验证。结果(1)训练集“切除肝段数”项与验证集临床资料比较,差别有统计学意义(P<0.05),其他项目差别无统计学意义。(2)手术时间预测模型训练集、验证集与真实值差别的RMSE、MAPE分别为(42.73 min,18.3%)和(47.11 min,27.4%)。(3)虽然训练集及内部验证中模型A较模型B有更高的预测准确率(0.929 vs 0.826/0.908 vs 0.854),更低的Brier评分(0.073 vs 0.109/0.088 vs 0.118),更好的预测性能(AUC:0.980 vs 0.900/0.942 vs 0.879),但模型B仍拥有较好的准确性及预测性能;在验证集中模型B与模型A有一致的准确率(0.824)、Brier评分(0.149),AUC差别无统计学意义(0.790 vs 0.770,P>0.05)。模型B与模型A在训练集、验证集及内部验证的预测结果有中~高度的一致性(Kappa为0.569~0.911)。bootstrap结果显示模型A、B表现稳定。结论利用术前资料联合MDN、RF算法有助于构建良好性能的肝癌切除术中出血预测模型,但需要多中心大样本数据进一步验证。

Objective To explore the feasibility and application value of building a prediction model for intraoperative bleeding in liver cancer resection with the help of machine learning algorithm using preoperative data.Methods The clinical data of 336 patients with primary liver cancer were retrospectively analyzed,and 268 cases in the training set and 68 cases in the test set were randomly selected by the ratio of 8∶2.Model A and model B was established respectively by random forest(RF)model based on actual-operation-time or based on predicted-operation-time conducted by mixture-density-network(MDN)model.The univariate regression analysis of dichotomous outcome was conducted by using logistic regression model.Variables of continuous-outcome prediction model were screened by Lasso regression,then the accuracy of the model measured by root mean square error(RMSE)and average absolute percentage error(MAPE).Classification prediction model was evaluated by accuracy,calibration curve,Brier score and area under ROC curve(AUC).The model performance stability was verified by the bootstrap method of all samples.Results(1)There was significant difference in the item of"number of hepatectomy segments"between training set and validation set(P<0.05)while no significant difference for other items.(2)The RMSE and MAPE value of operation-time prediction model were 42.73 min and 18.3%in the training set,while 47.11 min and 27.4%in the validation set.(3)Model A in the training set and bootsrtap validation had higher prediction accuracy(0.929 vs 0.826/0.908 vs 0.854),a lower Brier score(0.073 vs 0.109/0.088 vs 0.118),and better prediction performance(AUC:0.980 vs 0.900/0.942 vs 0.879)than model B,which still performed satisfactorily.In the validation set,model B and model A had the same accuracy rate(0.824)and Brier score(0.149),and there was no significant difference in AUC(0.790 vs 0.770,P>0.05).Model B and Model A had medium to high degree of agreement between the predictions of the training set,the validation set,and the bootstrap validation(Kappa 0.569-0.911).Meanwhile,the bootstrap validation showed that model A and B performed stably.Conclusion Combining MDN and RF algorithms with preoperative data was helpful to develop a good predictive model for intraoperative bleeding in liver cancer resection,which requires further verification with multi-center large sample data.