基于机器学习算法的前列腺癌诊断模型研究

Diagnostic Model Research of Prostate Cancer Based on Machine Learning Algorithm

目的基于机器学习的3种算法建立诊断预测模型,比较3种模型对前列腺癌的诊断价值。方法选择2008~2014年在中国人民解放军总医院进行前列腺穿刺活检的患者956例（其中前列腺癌463例,前列腺增生493例）,采用Logistic回归分析,筛选出预测因子（年龄、游离之前列腺特异抗原、游离之前列腺特异抗原百分比、前列腺体积和前列腺特异性抗原密度）。应用基于机器学习的BP神经网络、Logistic回归和随机森林算法构建诊断预测模型,比较3种模型对前列腺癌的预测准确性。结果 Logistic回归、BP神经网络和随机森林模型对前列腺癌的诊断能力比任一单项指标都高,3种模型的灵敏度分别为77.5%、77.4%、76.2%,特异度分别为74.8%、76.8%、76.9%,精确度分别为76%、77%、77%,受试者工作特征曲线下面积（AUC）分别为0.831、0.832、0.833,3种模型对前列腺癌的诊断能力没有显著性差异。结论上述结果验证了3种模型均具有较高的诊断有效性,可将模型纳入泌尿决策,协助临床医生对前列腺癌患者进行诊断和治疗,并减少不必要的活检。

Objective To establish diagnostic prediction models based on three machine learning algorithms and compare the value of the three models in the diagnosis of prostate cancer（PC）. Methods The research selected the clinical data of 956 patients（including 463 cases of prostate cancer and 493 cases of benign prostatic hyperplasia） with prostate biopsy in the General Hospital of PLA during 2008~2014. Predictors were screened by Logistic regression which included age, free prostate-specific antigen（f PSA）, the percentage of free prostate-specific antigen（free PSA/total PSA）, prostate volume, and PSA density（PSAD）. The paper further compared the diagnostic accuracy of three models in the prediction of prostate cancer by using BP neural network, Logistic regression（LR）, and random forest algorithm based on machine learning. Results The diagnostic capability of Logistic regression, BP neural networks, and random forest model for prostate cancer was higher than any a single indicator. Retrospectively, the sensitivity of the three models were 77.5%, 77.4%, and 76.2%; the specificity was 74.8%, 76.8%, and 76.9%; the accuracy was 76%, 77%, and 77%. The area under the ROC curve（AUC） was 0.831 for LR model, 0.832 for BP neural networks model, and 0.833 for the random forest model respectively, which indicated that there were no statistically significant difference existing in the three modes in terms of diagnostic effectiveness. Conclusion The above results verified the high diagnostic validity of these three models, which all could be incorporated into urologic decision making to assist clinicians carry out diagnosis and treatment so as to reduce the unnecessary biopsies.