基于机器学习模型预测前列腺重复穿刺结果的临床研究

Clinical study on prediction of repeated prostate puncture results based on machine learning model

目的 基于机器学习算法建立、分析并筛选出理想的模型用于预测前列腺重复穿刺患者阳性率,以指导临床医师对患者重复穿刺的临床决策。方法 本研究回顾性分析自2008年1月至2022年12月期间来自多中心的281名连续接受至少两次以上前列腺活检且首次病理均为阴性的男性患者的临床资料。记录患者常规临床诊疗数据,包括BMI,两次穿刺记录的前列腺体积、PSA水平、fPSA/PSA比率、PSAD水平、穿刺后并发症、穿刺后病理学结果,以及两次前列腺穿刺采用的策略。统计分析比较最终病理阳性组与阴性组的参数差异情况。将参数分别通过多变量逻辑回归(LR)、K近邻搜索(KNN)、支持向量机(SVC)、决策树(DT)、随机森林分类器(RF)、朴素贝叶斯分类器(NBC)和梯度增强树(GB)7种机器学习算法模型进行训练拟合,将所有数据的70%作为训练集数据与30%的剩余验证集数据作6倍交叉验证,分析比较准确性和敏感性,采用受试者工作特征曲线评价各模型的诊断准确性,并最终评估其在预判前列腺癌第二次穿刺结果中的效果。结果 本组患者年龄32~87岁,平均69.03岁。最终穿刺病理中201例为阴性患者,80例确诊前列腺癌。组间比较显示第二次穿刺年龄、第二次穿刺前列腺体积、第二次穿刺PSA水平、第二次穿刺PSAD、第二次穿刺FPSA/PSA比率均有显著统计学差异(P<0.05)。将所有收集的数据随机分为两部分,70%(197例)前列腺穿刺患者用作训练数据集,剩余的30%(84例)患者用作验证数据集。结果显示在预测第二次前列腺穿刺结果方面,LR、KNN、SVC、DT、GNB、RF、GB的准确率波动于66.67%~74.88%;错误率波动于21.43%~34.29%;召回率波动于9.52%~47.62%;特异性波动于79.59%~97.96%;精确率波动于33.33%~80.00%。各模型的特异性分别为各个模型的ROC曲线下面积,波动于0.568~0.725。结论 SVC可以在较简单较少的参数下更准确地预测前列腺第二次穿刺的阳性情况。相对于其他算法模型,其使得预测结果拥有更好的灵敏性和特异性。通过更大样本量数据的收集和训练,其有潜力可以成为一项具有广泛配合度的预测前列腺第二次穿刺结果的检测工具。

Objective To establish,analyze,and screen ideal models based on machine learning algorithms to predict the positive rate of patients who received repeated prostate punctures for guiding clinical decision-making.Methods Clinical data of 281 male patients from multiple centers who underwent at least two consecutive prostate biopsies from January 2008 to December 2022,were retrospectively analyzed,and their first biopsy pathology result was recorded as negative.Key clinical diagnosis and treatment data were recorded for each patient,including BMI,prostate volume,PSA level,fPSA/PSA ratio,PSAD level,complications after the first puncture,pathological results after the first puncture,and strategies used for the two prostate punctures.Statistical analysis was used to analyze the diferences in parameters between the final puncture pathology positive group and the negative group.The parameters were trained and fited through seven machine learning algorithm models such as multivariable logical regression(LR),K-nearest neighbor search(KNN),support vector machine(SVC),decision tree(DT),random forest classifier(RF),naive Bayes classifier(NBC)and gradient enhancement tree(CB).70%of all data as training set data and 30%of the remaining as verification set data were used for six times cross validation to analyze and compare accuracy and sensitivity,the diagnostic accuracy of each model was evaluated using receiver operating characteristic curve,and its effectiveness in predicting the results of prostate cancer secondary puncture was ultimately evaluated.Results The patients were aged between 32 and 87 years old,with an average of 69.03 years old.Among the patients,201 patients were recorded as negative in the final puncture pathology,and 80 patients were recorded as prostate cancer.The inter-group analysis showed that there were significant statistical differences(P<0.05)in the age of second puncture,volume of first puncture prostate,volume of second puncture prostate,PSA level of second puncture,PSAD of second puncture,and FPSA/PSA ratio of second puncture.All collected data were randomly divided into two parts,such as the training dataset(197 cases,70%of prostate puncture patients)and the validation dataset(84 cases,the remaining 30%).The results showed that the accuracy rates of LR,KNN,SVC,DT,GNB,RF,and CB in predicting the results of secondary prostate puncture were found to fluctuate between 66.67%and 74.88%,the error rates to fluctuate between 21.43%and 34.29%,the recall rates to fluctuate between 9.52%and 47.62%,the specificity to fluctuate between 79.59 and 97.96%,the accuracy to fluctuate between 33.33%and 80.00%.The specificities of each model(the area under the ROC curve)were found to fluctuate between 0.568 and 0.725.Conclusion SVC can more accurately predict the positivity of secondary prostate puncture with fewer parameters.Compared to other algorithm models,it makes the prediction results to have better sensitivity and specificity.Through the collection and training of larger sample size data,it has the potential to become a detection tool with broad compatibility for predicting prostate secondary puncture results.