基于集成学习算法构建前列腺癌预测模型

目的：利用机器学习算法建立前列腺癌诊断预测模型，为前列腺癌患者的穿刺术前诊断提供参考。方法：收集2017年1月-2018年12月中国医科大学附属盛京医院泌尿外科接受前列腺穿刺的255例患者的临床信息作为变量，采用Logistic多因素分析、信息增益率两种方法筛选研究变量，应用十折交叉验证划分训练集和测试集，采用多种机器学习算法（RF，SVM，Logistic，Naive Bayes）建立前列腺癌诊断模型，收集2019年1-6月的75例患者作为验证集，进一步评估模型性能和临床应用的可能性。结果：应用信息增益率筛选变量所建立的模型性能优于Logistic多因素回归分析。在4种机器学习算法中，Naive Bayes算法AUC最高，在试验集和验证集上分别为0.826和0.797。RF算法的Precision最高，在试验集和验证集上分别达到0.839和0.791。结论：基于前列腺穿刺患者的多种临床信息，通过机器学习方法建立诊断预测模型具有较高的准确率，能够为前列腺癌的诊断提供一定参考。

Integrated learning algorithm-based establishment of prostate cancer prediction model

Objective To provide reference for the pre-puncture diagnosis of prostate cancer by establishing a prostate cancer prediction model using the machine learning algorithm. Methods The prostate cancer diagnostic model was established using RF, SVM, logistic and Naive Bayes machine learning algorithms with the clinical information of 255 prostate cancer patients(served as an experimental group) who underwent prostate puncture in our hospital from January 2017 to December 2019 served as its variables detected by multivariate logistic regression analysis and information gain rate analysis respectively. The performance and clinical application of the prostate cancer diagnostic model were further evaluated with 75 patients admitted to our hospital from January 2019 to June 2019 served as a validation group. Results The performance of the prostate cancer diagnostic model established with the variables detected by information gain rate analysis was better than that established with the variables detected by multivariate logistic regression analysis. The AUC measured by Naive Bayes algorithm was larger than that measured by RF, SVM and logistic algorithms, which was 0.826 and 0797 respectively in experimental group and validation group. The accuracy of RF algorithm was higher than that of naive Bayes, SVM and logistic algorithms, which was 0.839 and 0.791 respectively in experimental group and validation group. Conclusion The accuracy of prostate cancer diagnostic model established using the machine learning algorithms based on the clinical information of prostate cancer patients is rather high, and can thus provide certain reference for the diagnosis of prostate cancer.