Abstract：
OBJECTIVE: The objective was to develop a prediction model for urinary tract infection (UTI) after pelvic surgery.
METHODS: We utilized data from three tertiary care centers of women undergoing pelvic surgery. The primary outcome was a UTI within 8 weeks of surgery. Additional variables collected included procedural data, severity of prolapse, use of mesh, anti-incontinence surgery, EBL, diabetes, steroid use, estrogen use, postoperative catheter use, PVR, history of recurrent UTI, operative time, comorbidities, and postoperative morbidity including venous thromboembolism, surgical site infection. Two datasets were used for internal validation, whereas a third dataset was used for external validation. Algorithms that tested included the following: multivariable logistic regression, decision trees (DTs), naive Bayes (NB), random forest (RF), gradient boosting (GB), and multilayer perceptron (MP).
RESULTS: For the training dataset, containing both University of British Columbia and Mayo Clinic Rochester data, there were 1,657 patients, with 172 (10.4%) UTIs; whereas for the University of Calgary external validation data, there were a total of 392 patients with a UTI rate of 16.1% (n = 63). All models performed well; however, the GB, DT, and RF models all had an area under the curve (AUC) > 0.97. With external validation the model retained high discriminatory ability, DT: AUC = 0.88, RF: AUC = 0.88, and GB: AUC = 0.90.
CONCLUSIONS: A model with high discriminatory ability can predict UTI within 8 weeks of pelvic surgery. Future studies should focus on prospective validation and application of randomized trial models to test the utility of this model in the prevention of postoperative UTI.
METHODS: We utilized data from three tertiary care centers of women undergoing pelvic surgery. The primary outcome was a UTI within 8 weeks of surgery. Additional variables collected included procedural data, severity of prolapse, use of mesh, anti-incontinence surgery, EBL, diabetes, steroid use, estrogen use, postoperative catheter use, PVR, history of recurrent UTI, operative time, comorbidities, and postoperative morbidity including venous thromboembolism, surgical site infection. Two datasets were used for internal validation, whereas a third dataset was used for external validation. Algorithms that tested included the following: multivariable logistic regression, decision trees (DTs), naive Bayes (NB), random forest (RF), gradient boosting (GB), and multilayer perceptron (MP).
RESULTS: For the training dataset, containing both University of British Columbia and Mayo Clinic Rochester data, there were 1,657 patients, with 172 (10.4%) UTIs; whereas for the University of Calgary external validation data, there were a total of 392 patients with a UTI rate of 16.1% (n = 63). All models performed well; however, the GB, DT, and RF models all had an area under the curve (AUC) > 0.97. With external validation the model retained high discriminatory ability, DT: AUC = 0.88, RF: AUC = 0.88, and GB: AUC = 0.90.
CONCLUSIONS: A model with high discriminatory ability can predict UTI within 8 weeks of pelvic surgery. Future studies should focus on prospective validation and application of randomized trial models to test the utility of this model in the prevention of postoperative UTI.
摘要：
目的：目的是建立盆腔手术后尿路感染（UTI）的预测模型。
方法：我们利用了三个三级护理中心的盆腔手术患者的数据。主要结果是手术后8周内的UTI。收集的其他变量包括程序数据，脱垂的严重程度，使用网格，防失禁手术,EBL,糖尿病,使用类固醇,雌激素的使用，术后使用导管,PVR,复发性尿路感染的病史，手术时间,合并症,和术后发病率，包括静脉血栓栓塞，手术部位感染。两个数据集用于内部验证，而第三个数据集用于外部验证。测试的算法包括以下内容：多变量逻辑回归，决策树(DTs)，朴素贝叶斯(NB)，随机森林(RF)，梯度增强(GB)，和多层感知器（MP）。
结果：对于训练数据集，包含不列颠哥伦比亚大学和罗切斯特梅奥诊所的数据，有1657名患者,172(10.4%)UTI；而对于卡尔加里大学的外部验证数据，共有392例患者,UTI发生率为16.1%(n=63).所有模型都表现良好；然而，GB,DT,和RF模型均具有曲线下面积(AUC)>0.97。通过外部验证，该模型保留了较高的判别能力，DT:AUC=0.88,RF:AUC=0.88,GB:AUC=0.90。
结论：具有高辨别能力的模型可以预测盆腔手术8周内的UTI。未来的研究应集中在前瞻性验证和随机试验模型的应用上，以测试该模型在预防术后UTI中的实用性。
方法：我们利用了三个三级护理中心的盆腔手术患者的数据。主要结果是手术后8周内的UTI。收集的其他变量包括程序数据，脱垂的严重程度，使用网格，防失禁手术,EBL,糖尿病,使用类固醇,雌激素的使用，术后使用导管,PVR,复发性尿路感染的病史，手术时间,合并症,和术后发病率，包括静脉血栓栓塞，手术部位感染。两个数据集用于内部验证，而第三个数据集用于外部验证。测试的算法包括以下内容：多变量逻辑回归，决策树(DTs)，朴素贝叶斯(NB)，随机森林(RF)，梯度增强(GB)，和多层感知器（MP）。
结果：对于训练数据集，包含不列颠哥伦比亚大学和罗切斯特梅奥诊所的数据，有1657名患者,172(10.4%)UTI；而对于卡尔加里大学的外部验证数据，共有392例患者,UTI发生率为16.1%(n=63).所有模型都表现良好；然而，GB,DT,和RF模型均具有曲线下面积(AUC)>0.97。通过外部验证，该模型保留了较高的判别能力，DT:AUC=0.88,RF:AUC=0.88,GB:AUC=0.90。
结论：具有高辨别能力的模型可以预测盆腔手术8周内的UTI。未来的研究应集中在前瞻性验证和随机试验模型的应用上，以测试该模型在预防术后UTI中的实用性。
