基于机器学习的宫颈癌根治术后尿潴留风险预测模型构建

Prediction model of uroschesis rate after radical cervical cancer resection based on machine learning

目的基于机器学习构建宫颈癌根治术患者尿潴留风险预测模型,并进行内部验证评价模型的预测效果,以期为宫颈癌根治术患者尿潴留的早期防治提供参考。方法便利抽样法选取2017年6月至2022年2月在安徽医科大学第一附属医院内实施宫颈癌根治手术的981例患者,按7∶3的比例分为训练集(687例)和测试集(294例)。通过文献回顾和危险因素分析,探索宫颈癌根治术后发生尿潴留的影响因素,采用机器学习中XGBoost、随机森林、支持向量机和决策树4种方法构建尿潴留风险预测模型。采用十折交叉验证的方式计算4种机器学习算法的精确率、召回率、F1值和AUC,筛选出预测效能最高的模型。结果纳入的981例患者中,宫颈癌根治术后尿潴留发生率为18.86%(185/981),尿潴留组中位年龄为51岁,非尿潴留组为50岁。将单因素分析中有统计学意义的变量和文献回顾总结的影响因素进行特征提取,纳入患者年龄、术中出血量、BMI、癌症分期、手术方式、手术切除范围、是否行盆腔淋巴结清扫术、合并症和残余尿的情况。在机器学习的4种模型构建方法中,随机森林模型效果最好,其训练集F1值为0.94,测试集F1值为0.77,绘制ROC并计算AUC为0.73。年龄、术中出血量、BMI、癌症分期、手术方式等指标对随机森林模型分类贡献较高。结论基于随机森林方法建立宫颈癌根治术后尿潴留风险预测模型效能最佳,该模型有助于护理工作人员评估患者发生尿潴留的风险,进而采取有针对性的护理干预积极预防术后尿潴留。

Objective To construct a risk prediction model for urinary retention in patients undergoing radical cervical cancer surgery based on machine learning,and the prediction effect of the model was internally verified and evaluated,in order to provide reference for the early prevention and treatment of urinary retention in patients undergoing radical cervical cancer surgery.Methods A total of 981 patients who underwent radical cervical cancer surgery in the First Affiliated Hospital of Anhui Medical University from June 2017 to February 2022 were selected and divided into the training set(687 cases)and the test set(294 cases)according to a ratio of 7∶3.Through literature review and risk factor analysis,the influencing factors of urinary retention after radical treatment of cervical cancer were explored,and the risk prediction model of urinary retention was constructed by using XGBoost,random forest,support vector machine and decision tree in machine learning.The accuracy rate,recall rate,F1 value and AUC of four machine learning algorithms were calculated by using the method of 10-fold cross-validation,and the model with the highest predictive efficiency was selected.Results Among the 981 patients included,the incidence of urinary retention after radical cervical cancer surgery was 18.86%(185/981).The median age of urinary retention group was 51 years old,and that of non urinary retention group was 50 years old.Statistically significant variables in the univariate analysis and influencing factors summarized by literature review were featured,including patient age,intraoperative blood loss,body mass index(BMI),cancer stage,surgical method,surgical resection scope,whether pelvic lymph node dissection was performed,comorbidities and residual urine.Among the four model building methods of machine learning,the random forest model has the best effect,its training set F1 value was 0.94,the test set F1 value was 0.77,the ROC was plotted and the AUC was calculated to be 0.73.Age,intraoperative blood loss,BMI,cancer stage and surgical method contributed significantly to the classification of random forest model.Conclusions The prediction model of urinary retention risk after radical cervical cancer surgery based on random forest method has the best efficacy.It is useful to help nursing personnel evaluate the risk of the uroschesis for a patient and then take targeted nursing interventions to actively prevent postoperative urinary retention.