基于决策树C5.0和Logistic回归构建前列腺癌根治术后尿失禁预测模型的比较研究

Comparative study on urinary incontinence prediction model constructed after radical prostatectomy based on decision tree C5.0 and Logistic regression

目的采用决策树C5.0与Logistic回归分别构建前列腺癌根治术后发生尿失禁的风险预测模型,对比2种模型的预测效果。方法采用便利抽样法,选取2019年1月—2022年1月在海军军医大学第一附属医院行前列腺癌根治术的前列腺癌患者260例。回顾性收集患者的临床资料,根据患者术后随访半年期间是否发生尿失禁进行分组,分为尿失禁组(n=74)和无尿失禁组(n=186)。比较两组患者的临床资料,采用决策树C5.0和Logistic回归分别建立前列腺癌根治术后尿失禁的风险预测模型。使用受试者工作特征(ROC)曲线、阴性预测值、阳性预测值、准确度、灵敏度、特异度和约登指数检验2种模型的预测性能。收集同期在海军军医大学第一附属医院行前列腺癌根治术的150例前列腺癌患者,分别对2种预测模型进行外部验证。结果决策树C5.0和Logistic回归模型均显示,术前前列腺体积≥40 ml、年龄≥60岁均为前列腺癌根治术后尿失禁的独立危险因素(P<0.05),术中保留完整膀胱颈是前列腺癌根治术后尿失禁的保护性因素(P<0.05);同时,Logistic回归分析还显示体重指数≥24 kg/m^(2)为术后尿失禁发生的危险因素(P<0.05)。决策树C5.0与Logistic回归模型的准确率分别为75.0%、74.6%;阳性预测值分别为58.2%、59.1%;阴性预测值分别为71.5%、77.8%;灵敏度分别为71.6%、68.9%;特异度分别为67.2%、74.7%;约登指数分别为38.8%、43.6%;ROC曲线下面积分别为0.744、0.777。外部验证结果显示,决策树C5.0模型的ROC曲线下面积为0.689,灵敏度和特异度分别为71.1%、66.7%,Logistic回归模型的ROC曲线下面积为0.762,灵敏度和特异度分别为77.8%、63.8%。结论年龄≥60岁、体重指数≥24 kg/m 2、术前前列腺体积≥40 ml为前列腺癌根治术后尿失禁发生的独立危险因素,而术中保留完整膀胱颈会降低术后尿失禁的发生风险。根据上述因素构建的Logistic回归模型对前列腺癌根治术后尿失禁的预测性能优于决策树C5.0模型。

ObjectiveTo construct risk prediction models for postoperative urinary incontinence in patients undergoing radical prostatectomy using decision tree C5.0 and Logistic regression,and to compare the predictive effects of the two models.MethodsA total of 260 patients with prostate cancer who underwent radical prostatectomy in the First Affiliated Hospital of Naval Medical University of PLA from January 2019 to January 2022 were selected by the convenient sampling method.Clinical data of the patients were retrospectively collected,and the patients were divided into the urinary incontinence group(n=74)and the non-urinary incontinence group(n=186)according to whether they had urinary incontinence during the six-month follow-up.The clinical data of the two groups were compared.Decision tree C5.0 and Logistic regression were retrospectively used to establish the risk prediction model of urinary incontinence after radical prostatectomy.The predictive performance of the two models was tested using receiver operating characteristic(ROC)curve,negative predictive value,positive predictive value,accuracy,sensitivity,specificity and Youden index.A total of 150 patients with prostate cancer undergoing radical prostatectomy in the First Affiliated Hospital of Naval Medical University of PLA were collected and the two prediction models were conducted external validation.ResultsBoth decision tree C5.0 and Logistic regression models showed that preoperative prostate volume≥40 ml and age≥60 years old were independent risk factors for postoperative urinary incontinence in patients undergoing radical prostatectomy(P<0.05),while preserving the intact bladder neck during surgery was a protective factor for postoperative urinary incontinence in patients undergoing radical prostatectomy(P<0.05).Meanwhile,Logistic regression analysis also showed that the body mass index≥24 kg/m^(2) was a risk factor for postoperative urinary incontinence(P<0.05).The accuracy of decision tree C5.0 and Logistic regression models was 75.0%and 74.6%,respectively.The positive predictive values were respectively 58.2%and 59.1%,and negative predictive values were respectively 71.5%and 77.8%.The sensitivity was respectively 71.6%and 68.9%,and the specificity was respectively 67.2%and 74.7%.The Youden index was 38.8%and 43.6%,respectively.The areas under ROC curve were 0.744 and 0.777,respectively.External verification results showed that the area under ROC curve of the Decision Tree C5.0 model is 0.689,with a sensitivity and specificity of 71.1%and 66.7%,respectively,and the area under the Logistic regression model was 0.762,and the sensitivity and specificity were 77.8%and 63.8%,respectively.ConclusionsAge≥60 years old,body mass index≥24 kg/m 2 and preoperative prostate volume≥40 ml are independent risk factors for urinary incontinence after radical prostatectomy,while maintaining the intact bladder neck during surgery reduce the risk of postoperative urinary incontinence.The Logistic regression model constructed based on the above factors has better predictive performance for urinary incontinence after radical prostatectomy than the decision tree C5.0 model.