探讨基于不同的机器学习技术构建肝移植术后发生早期严重并发症预测模型的效能。

回顾性分析2018年7月至2024年8月在南方医科大学南方医院接受原位肝移植的129例患者临床资料。患者或家属签署知情同意书。其中男121例，女8例；年龄18~71岁，中位年龄51岁。根据Clavien-Dindo分级系统，将肝移植术后90 d内是否发生Ⅲb级及以上并发症的患者分为严重并发症组和非严重并发症组。使用归一化法对连续变量进行预处理，应用信息增益、Boruta算法和弹性网络分别做特征选择，纳入这些方法得到的结果中共有的因素，分别采取支持向量机、随机森林（RF）、逻辑回归、极端梯度提升和朴素贝叶斯5种机器学习技术，构建肝移植术后发生早期严重并发症的预测模型。采取十折交叉验证的方式计算不同机器学习算法的准确率、精确率、召回率、F1值和ROC AUC。在最优模型中使用沙普利可加性特征解释（SHAP）法评估预测模型中各特征因子的相对重要性。

肝移植术后早期严重并发症的发生率为24.8%（32/129）。纳入术前控制营养状态（COUNT）评分、术前腰大肌厚度/身高值（PMTH）、术前MELD评分、手术无肝期时间和术中失血量5个特征因子输入机器学习模型。综合各项指标，RF预测模型平均效能较好，在召回率（0.844）和F1分数（0.866）上表现最佳，显示出其在实际应用中的优秀能力，在验证集中AUC、准确度、精确率分别为0.808、0.800、0.898。RF预测模型预测变量重要性可视化排序依次为术前COUNT评分、手术无肝期、术前MELD评分、术中失血量、PMTH。根据SHAP汇总图，术前COUNT评分、术中失血量、手术无肝期时间、术前MELD评分与术后早期严重并发症的发生呈正相关，而与术前PTMH呈负相关。

本研究基于不同机器学习技术构建了多个肝移植术后发生早期严重并发症的预测模型，其中RF预测模型具有良好的预测性能，对探索肝移植患者围手术期管理具有一定意义。

To evaluate the efficiency of prediction models constructed based on different machine learning technologies for early severe complications after liver transplantation.

Clinical data of 129 patients undergoing orthotopic liver transplantation in Nanfang Hospital of Southern Medical University from July 2018 to August 2024 were retrospectively analyzed. The informed consents of all patients or family members were obtained. Among them, 121 patients were male and 8 female, aged from 18 to 71 years, with a median age of 51 years. According to Clavien-Dindo classification, all patients were divided into the severe and non-severe complication groups according to the incidence of Clavien-Dindo grade Ⅲb or above complications within 90 d after liver transplantation. Normalization method was used to preprocess continuous variables. Information gain, Boruta and elastic network were utilized for feature selection. Common factors in the methods and results were included. Five machine learning technologies, including support vector machine (SVM), random forest (RF), logistic regression (LR), eXtreme Gradient Boosting (XGBoost) and naive Bayes, were adopted to construct prediction models for early severe complications after liver transplantation. The accuracy rate, precision rate, recall rate, F1 value and the area under the ROC curve (AUC) of different machine learning algorithms were calculated by ten-fold cross-validation. Shapley additive explanations (SHAP) was adopted in the optimal model to evaluate the relative importance of each feature in the prediction models.

The incidence of early severe complications after liver transplantation was 24.8%(32/129). Five features including preoperative controlling nutritional status (COUNT) score, preoperative psoas muscle thickness per height (PMTH), preoperative MELD score, anhepatic phase and intraoperative blood loss, were input into machine learning models. Considering all indexes, the average efficiency of the RF prediction model was high and yielded the highest recall rate (0.844) and F1 score (0.866), demonstrating excellent capability in practical application. The average AUC, accuracy and precision rates in the validation set were 0.808, 0.800 and 0.898, respectively. The visual ranking of the importance of prediction variables in RF prediction models was: COUNT score, anhepatic phase, preoperative MELD score, intraoperative blood loss and PMTH, respectively. According to the SHAP, preoperative COUNT score, intraoperative blood loss, anhepatic phase and preoperative MELD score were positively correlated with, while preoperative PTMH value was negatively correlated with the incidence of early severe complications after liver transplantation.

Prediction models for early severe complications after liver transplantation are constructed based on different machine learning technologies. RF prediction model has high prediction performance, which is of certain significance to investigate perioperative management of liver transplantation recipients.