基于机器学习预测前列腺增生术后早期生化复发

Prediction of Early Biochemical Recurrence after Prostatic Hyperplasia Surgery Based on Machine Learning

摘要: 前列腺增生的术后生化复发是困扰患者正常生活的难题之一，其中尿失禁是最突出的问题。传统的统计学方法和逻辑回归算法在对术后尿失禁的预测上具有很大的局限性，预测效果也不尽人意，大大影响了对患者术后的治疗和术前的诊断。对于以上的问题，文章构建了新的预测前列腺增生术后早期生化复发的模型，主要目的是更准确、更有效地预测患者术后早期生化复发，将该模型命名为“S-KFG”。在研究中，模型主要是以软投票分类器作为集成算法，将KNN、随机森林、梯度提升树组合形成。另外使用堆叠算法，以随机森林算法为基模型、逻辑回归算法为元模型做集成模型，命名为“Hancerforest”。并与一些传统的机器学习算法，如随机森林、梯度提升树等，以及传统的逻辑回归预测进行比较。基于大连医科大学附属第二医院泌尿外科的数据，S-KFG模型的结果显示：模型的AUC值达到了0.888，Hancerforest模型的AUC值也达到了0.872，与传统的逻辑回归预测模型相比，本研究的模型具有更好的预测效果。

Abstract: Postoperative biochemical recurrence of prostatic hyperplasia is one of the problems that plague the normal life of patients, among which urinary incontinence is the most prominent problem. Traditional statistical methods and logistic regression algorithms have great limitations in predicting postoperative urinary incontinence, and the prediction effect is unsatisfactory, which greatly affects the postoperative treatment and preoperative diagnosis of patients. In order to solve the above problems, a new model for predicting early biochemical recurrence after prostatic hyperplasia was constructed, and the main purpose was to predict early biochemical recurrence in patients with prostatic hyperplasia more accurately and effectively, named “S-KFG”. In this paper, the model mainly uses the soft vote classifier as the ensemble algorithm and combines KNN, random forest, and gradient-boosting trees to form a group. In addition, the stacking algorithm is used, the random forest algorithm is used as the base model, and the logistic regression algorithm is used as the metamodel to make the ensemble model, which is named “Hancerforest”. It is also compared with some traditional machine learning algorithms such as random forest, gradient boosted tree, etc., as well as traditional logistic regression prediction. Based on the data of the Department of Urology, the Second Affiliated Hospital of Dalian Medical University, the results of the S-KFG model showed that the AUC value of the model reached 0.888, and the AUC value of the Hancerforest model also reached 0.872, which was better than the traditional logistic regression prediction model.