基于机器学习算法构建食道闭锁术后吻合口漏概率在线交互式网页计算工具及相应风险分层系统

Construction of an online interactive calculation tool and corresponding risk stratification system for the probability of postoperative anastomotic leak in esophageal atresia based on machine learning algorithms

摘要

目的 利用机器学习技术对食道闭锁术后出现吻合口漏进行预测，寻找导致术后出现吻合口漏的危险因素，计算相应截断值，并制作交互式网页计算工具，方便医务人员快速计算术后出现吻合口漏的具体风险概率。 方法 收集2009年1月至2021年12月在重庆医科大学附属儿童医院胸心外科接受手术治疗的251例Ⅲ型先天性食道闭锁患者的临床资料。包括患儿人口学特征、手术资料和术后资料。本课题组采用支持向量机（support vector machine，SVM）、随机森林（random forest，RF）、逻辑回归模型（logistic regression，LR）、XGboost分类（eXtreme gradient boosting，XGBoost）、高斯朴素贝叶斯（gaussian naive bayes，GNB）这5种机器学习算法来构建预测食道闭锁术后吻合口漏的预测模型。利用受试者工作特征（receiver operating characteristic，ROC）曲线的曲线下面积（area under thecurve，AUC）评价效度，同时综合F1分数、准确率、灵敏度及特异度，Hosmer-Lemeshow检验Brier分数评价校准度及临床决策曲线（decision curve analysis，DCA曲线）对模型的校准度及稳定性进行评价。利用限制性立方样条分别计算相应危险因素的截断值，最后制作交互式网页计算工具，构建术后吻合口漏风险分层系统，方便医务人员快速使用。 结果 通过对候选风险因素进行单因素分析、重要度排序、LASSO回归（least absolute shrinkage and selection operator，LASSO）筛选出危险因素为断端距离、是否合并复杂先心、术前蛋白、是否合并肺部感染。在5种机器学习算法中，逻辑回归模型在ROC曲线和DCA性能及校准曲线综合指标方面表现最佳，在逻辑回归模型中，训练集的AUC为0.828，准确度为0.772，F1分数为0.532，验证集的AUC为0.799，准确度为0.765，F1分数为0.544。提示该模型用于预测Ⅲ型先天性食道闭锁术后出现吻合口漏有较好的区分度及校准度。利用限制性立方样条，计算了断端距离及术前蛋白的截断值分别为2 cm及33.9 g/L，临床医务人员可以利用在线交互式网页计算工具，输入相应危险因素的结果，计算出某1个患者术后出现吻合口漏的具体概率值。 结论 逻辑回归模型可较好地预测Ⅲ型先天性食道闭锁患儿术后出现吻合口漏危险因素，在线交互式网页计算工具可以迅速地计算出术后吻合口漏的概率，方便医务人员使用。

Abstract

Objective To predict postoperative anastomotic leak in esophageal atresia using machine learning techniques，to identify the risk factors for postoperative anastomotic leak，to calculate corresponding cut-off values，to develop an interactive web-based tool，and to help healthcare professionals quickly calculate the specific risk probability of postoperative anastomotic leak. Methods Clinical data were collected from 251 patients with type Ⅲ congenital esophageal atresia who underwent surgical treatment in our hospital from January 2009 to December 2021，including demographic features，surgical data，and postoperative data. Five machine learning algorithms，i.e.，support vector machine（SVM），random forest （RF），logistic regression（LR），XGBoost，and Gaussian naive Bayes（GNB），were used to construct a predictive model for anastomotic leak after esophageal atresia repair. The area under the ROC curve（AUC），F1 score，accuracy，sensitivity，and specificity were used to evaluate the validity of the model，the Hosmer-Lemeshow test and Brier score were used to evaluate the degree of calibration，and the decision curve analysis （DCA curve） was used to evaluate the degree of calibration and stability. Restricted cubic spline techniques were used to calculate the cut-off value of each risk factor，and then an interactive web-based calculation tool was developed to establish a risk stratification system for postoperative anastomotic leak，which was used to facilitate healthcare professionals in convenient application. Results The univariate analysis，importance ranking，and LASSO regression were performed for candidate risk factors，and the results showed that the distance between the ends of the esophageal gap，presence or absence of complex congenital heart disease，preoperative protein level，and presence or absence of pulmonary infection were the risk factors for postoperative anastomotic leak. Among the five machine learning algorithms，the logistic regression model exhibited the best performance in terms of AUC，DCA，and calibration curve，with an AUC of 0.828，an accuracy of 0.772，and an F1 score of 0.532 in the training set and an AUC of 0.799，an accuracy of 0.765，and an F1 score of 0.544 in the validation set，suggesting that the model had good discriminatory ability and degree of calibration in predicting postoperative anastomotic leak in type Ⅲcongenital esophageal atresia. Meanwhile，the restricted cubic spline analysis showed that the distance between the ends of the esophageal gap and preoperative protein level had a cut-off value of 2 cm and 33.9 g/L，respectively，and healthcare professionals could use the online interactive web-based tool to input the results of related risk factors and calculate the specific probability of postoperative anastomotic leak for a given patient. Conclusion The logistic regression model can predict the risk factors for postoperative anastomotic leak in patients with type Ⅲ congenital esophageal atresia，and the online interactive web-based tool is designed to quickly calculate the probability of postoperative anastomotic leak，thereby providing convenience for healthcare professionals.

关键词

食道闭锁 / 机器学习 / 预测

Key words

esophageal atresia / machine learning / prediction