【摘要】 目的 探讨基于术前临床指标构建的机器学习模型在预测肝部分切除术后低白蛋白血症发生风险中的应用价值。方法 采用回顾性研究方法，收集2018年1月—2023年1月于南方医科大学南方医院行肝部分切除术的857例患者的临床资料，包括人口统计学资料、基础疾病史、肿瘤特征、术前实验室指标和围手术期指标。按7∶3比例将研究数据划分为训练集与测试集。正态分布的计量资料2组间比较采用成组t检验；偏态分布计量资料2组间比较采用两独立样本的Wilcoxon秩和检验；计数资料2组间比较采用χ~2检验或Fisher确切概率法。应用最小绝对收缩和选择算子（LASSO）回归筛选特征变量，采用逻辑回归、决策树、人工神经网络、K近邻（KNN）、支持向量机、极度梯度提升和轻量级梯度提升机共7种机器学习算法构建预测模型。采用受试者操作特征曲线下面积（AUC）评估模型区分度，采用DeLong检验比较模型的AUC。采用校准曲线与决策曲线分析评估模型的校准度和临床实用性，并与白蛋白-胆红素（ALBI）评分、终末期肝病模型（MELD）评分进行比较，并利用沙普利加性解释法（SHAP）解释最佳模型的关键影响因素。结果 最终纳入700例患者，术后发生低白蛋白血症患者283例（40.42%）。LASSO回归筛选出年龄、乙型肝炎、脂肪肝、阻断时间、术前白蛋白（Alb）、手术时长、术中出血量和术前天冬氨酸氨基转移酶（AST）共8个预测因子。在7种机器学习模型中，KNN模型的综合预测效能最优，测试集的AUC为0.835(95%CI:0.781～0.889)，敏感度为84.0%，特异度为65.5%。ALBI评分和MELD评分的AUC分别为0.652和0.524,KNN模型的预测效能显著优于上述两项评分（Z值分别为5.309、8.945,P值均<0.001）。校准曲线显示，KNN模型的预测概率与实际发生率一致性良好；决策曲线分析表明，KNN模型在较大阈值范围内具有临床净收益。SHAP分析显示，术前Alb、乙型肝炎、手术时长、年龄为更显著的影响因素，且乙型肝炎与年龄及手术时长存在协同效应。结论 基于术前临床指标构建的KNN机器学习模型可有效预测肝部分切除术后低白蛋白血症的发生风险，性能优于传统模型，可为临床早期识别高风险患者提供参考依据。更多还原

【Abstract】 Objective To investigate the application value of a machine learning model based on preoperative clinical indicators in predicting the risk of hypoalbuminemia after partial hepatectomy. Methods A retrospective analysis was performed for the clinical data of 700 patients who underwent partial hepatectomy in Nanfang Hospital, Southern Medical University, from January 2018 to January 2023, including demographic data, history of underlying diseases, tumor characteristics, preoperative laboratory markers, and perioperative indicators. The research data were divided into a training set and a test set at a ratio of 7∶3. The twoindependent-samples t test was used for comparison of normally distributed continuous data between two groups; the twoindependent-samples Wilcoxon rank-sum test was used for comparison of continuous data with skewed distribution between two groups; the chi-square test or the Fisher’s exact test was used for comparison of categorical data between two groups. The least absolute shrinkage and selection operator(LASSO) regression analysis was used to identify characteristic variables, and 7 machine learning algorithms were used to construct predictive models, i.e., logistic regression, decision tree, artificial neural network, Knearest neighbors(KNN), support vector machine, eXtreme gradient boosting, and light gradient boosting machine. The receiver operating characteristic(ROC) curve and the area under the ROC curve(AUC) were used to assess the discriminatory ability of models, and the DeLong test was used for comparison of AUC. The calibration curve and decision curve analysis were used to assess the calibration and clinical practicability of models, and the models were compared with albumin-bilirubin(ALBI) score and Model for End-Stage Liver Disease(MELD) score. SHapley Additive exPlanations(SHAP) were used to interpret the key influencing factors for the optimal model. Results A total of 700 patients were finally enrolled, 283(40.42%) developed hypoalbuminemia after surgery. The LASSO regression analysis identified 8 predictive factors of age, hepatitis B, fatty liver, blockade time, preoperative albumin(Alb), time of operation, intraoperative blood loss, and preoperative aspartate aminotransferase(AST). Among the 7 machine learning models, the KNN model showed the best overall predictive performance, with an AUC of 0.835(95% confidence interval: 0.781 — 0.889), a sensitivity of 84.0%, and a specificity of 65.5% in the test set. ALBI and MELD scores had an AUC of 0.652 and 0.524, respectively, and the KNN model had a better predictive performance than these two scores(Z=5.309 and 8.945, both P <0.001). The calibration curve showed good consistency between predicted probabilities and actual incidence rates, and the decision curve analysis showed that the KNN model had net clinical benefit across a wide threshold range. The SHAP analysis showed that preoperative Alb, hepatitis B, time of operation, and age were the most significant influencing factors, and a synergistic effect was observed between hepatitis B and age/time of operation. Conclusion The KNN machine learning model constructed based on preoperative clinical indicators can effectively predict the risk of hypoalbuminemia after partial hepatectomy and has a better performance than traditional scoring models, which provides a reference for the early identification of high-risk patients in clinical practice.更多还原