全膝关节置换术后深静脉血栓预测模型的建立△

Establishment of the prediction models for the deep vein thrombosis after total knee arthroplasty

摘要:

［目的］ 构建基于逻辑回归和 XGBoost 算法的全膝关节置换术 （total knee arthroplasty, TKA） 围手术期深静脉血栓形成风险（deep vein thrombosis, DVT）的预测模型。［方法］回顾性分析 2017 年 12 月—2021 年 10 月于安徽医科大学附属安徽省立医院骨科接受 TKA 手术治疗的 3 711 例患者的临床资料，构建逻辑回归和 XGBoost 算法预测模型，筛选围手术期出现 DVT 的预测因素，并比较两者的预测效能。［结果］3 711 例患者中，TKA 术后共有 889 例患者发生 DVT，总发病率 23.96%。单项因素比较表明，与非 DVT 组相比，DVT 组年龄显著更大（P<0.05），低分子肝素和 X 因子抑制剂使用的比率显著更高（P<0.05）， 术后 6 h 早期抗凝比率显著更低（P<0.05）、术前准备时间显著更短（P<0.05）、术后住院时间显著更长（P<0.05），手术时间显著更长（P<0.05），术前 D-二聚体含量显著更高（P<0.05），术前凝血酶原活动度更低（P<0.05）、术后血磷含量更高（P<0.05）、术后尿素氮肌酐比值更高（P<0.05），差异均有统计学意义。逻辑回归表明：术后尿素氮肌酐比值升高（OR=1.576，P<0.05）、术后住院时间长（OR=1.393，P<0.05）、高龄（OR=1.214，P<0.05）、术后血磷升高（OR=1.160，P=0.05）、术前 D-二聚体升高（OR= 1.058，P=0.05）是发生 DVT 的危险因素。XGBoost 模型显示年龄、术后住院时间、术后 D-二聚体水平、血清尿素氮/肌酐比值、 使用低分子肝素是重要的特征向量。两种预测模型 ROC 分析的 AUC 分别为 0.709 和 0.840。［结论］XGBoost 模型对于 TKA 围手术期 DVT 事件具有良好的预测能力，患者年龄、术后住院时间、术后 D-二聚体含量、血清尿素氮/肌酐比值、使用低分子肝素是潜在的重要预测指标。

Abstract:

[Objective] To establish the prediction models of deep vein thrombosis (DVT) after total knee arthroplasty (TKA) based on the logic regression and the extreme gradient boosting (XGBoost) . [Methods] A retrospective study was conducted on 3 711 patients who re- ceived TKA in the Department of Orthopaedics, Anhui Provincial Hospital from December 2017 to October 2021. The prediction models for DVT after TKA were established based on the factors related the DVT by logical regression and XGBoost algorithm respectively, which were compared in term of prediction efficiency. [Results] Of 3 711 patients, 889 patients proved DVT after TKA, with a total incidence of 23.96%. In term of univariate comparison, the DVT group proved significantly older (P＜0.05) , higher ratio of low-grade heparin and X factor inhibi- tors used (P＜0.05) , lower early anti-coagulation 6 hours after the operation (P＜0.05) , shorter preoperative preparation time (P＜0.05) , lon- ger postoperative hospital stay (P＜0.05) , longer operation time (P＜0.05) , higher preoperative blood D-dimer (P＜0.05) , lower preoperative coagulation activity (P＜0.05) , higher level of postoperative blood phosphate (P＜0.05) and higher postoperative ratio of urea nitrogen to creati- nine (P＜0.05) than the non-DVT group. As results of logical regression, higher postoperative ratio of urea nitrogen to creatinine (OR=1.576, P＜0.05) , longer postoperative hospital stay (OR=1.393, P＜0.05) , older age (OR=1.214, P＜0.05) , higher postoperative serum phosphorus (OR=1.160, P=0.05) , higher preoperative D-dimer (OR=1.058, P=0.05) were risk factors for DVT events. In term of XGBoost model, the age, postoperative hospital stay, postoperative D-dimer, serum urea nitrogen/creatinine ratio, and use of low molecular weight heparin were important feature vectors. The AUC of ROC analysis of the two prediction models were of 0.709 and 0.840 respectively. [Conclusion] XG- Boost model has a good ability to predict DVT events in the perioperative period of TKA. The age, postoperative hospital stay, postoperative D-dimer, serum urea nitrogen/creatinine ratio, and low molecular weight heparin used are potential important predictors.