基于极限梯度提升算法的前列腺增生与前列腺癌鉴别诊断

Differential Diagnosis of BPH and Prostate Cancer Based onExtreme Gradient Boosting Algorithm

目的利用机器学习方法构建前列腺增生与前列腺癌的鉴别诊断模型,辅助临床医生进行穿刺术决策。方法利用中国人民解放军总医院第一医学中心2009年至2018年的前列腺增生与前列腺癌住院患者基本信息、既往史、临床信息等数据,分别利用Logistic回归和极限梯度(XGBoost)方法构建鉴别诊断模型,并利用灵敏度、特异性、准确率、AUC值和ROC曲线评估模型效果。结果Logistic回归和XGBoost方法构建的模型性能与单因素tPSA(P<0.001)、fPSA(P<0.001)、f/tPSA(P<0.001)分析差异均具有统计学意义。利用前列腺体积、f/tPSA、tPSA、fPSA、尿白细胞、无机磷等15个重要变量构建XGBoost模型,得到测试集的灵敏度、特异性、准确率、AUC值分别为0.835、0.815、0.826、0.903。结论多因素XGBoost模型较单因素(f/tPSA、tPSA、fPSA)预测模型和Logistic回归模型更优,具有更好的鉴别诊断能力,且对tPSA为4~10ng/mL的患者也具有一定的鉴别能力。

Objective To establish a classification model of prostatic hyperplasia and prostate cancer by using machine learning method,and to assist clinicians in the decision of puncture.Methods Logistic regression and eXtreme Gradient Boosting algorithm(XGBoost)were used to construct classification models based on the basic information,medical history and clinical information of inpatients with BPH and prostate cancer from 2009 to 2018 in the First Medical Center of the PLA General Hospital.Sensitivity,specificity,accuracy,AUC value and ROC curve analysis were used to evaluate the effectiveness of the models.Results Logistic regression and XGBoost methods were used to construct models.The performance and single factor tPSA(P<0.001),fPSA(P<0.001),f/tPSA(P<0.001)were significantly different.XGBoost model was constructed by 15 important variables including prostate volume,f/tPSA,tPSA,fPSA,urinary white blood cell and inorganic phosphorus,etc.The sensitivity,specificity,accuracy and AUC values of the test set were 0.835,0.815,0.826 and 0.903,respectively.Conclusion Compared with the single factor prediction model and logistic regression model,the constructed multi-factor XGBoost model has a better differential diagnosis ability,and has a certain differential ability for patients with tPSA of 4-10ng/mL.