基于机器学习模型分析不进行穿刺活检的前列腺根治术的可行性

目的基于机器学习算法构建前列腺癌（PCa）和临床显著前列腺癌（CSPCa）的诊断模型，并探讨不进行穿刺活检的前列腺根治术的可行性。 方法回顾性收集2016年4月至2020年3月期间符合纳入标准的688例前列腺穿刺患者的临床资料，包括年龄、PSA、前列腺体积、磁共振报告和穿刺活检病理结果，基于人工神经网络（ANN）、支持向量机（SVM）、随机森林（RF）和Logistic回归（LR）算法，构建PCa和CSPCa的诊断模型，采用受试者工作特征曲线评价各模型的诊断准确性，并分析不进行穿刺活检的前列腺根治术的可行性。 结果ANN（AUC=0.891和0.911）、SVM（AUC=0.903和0.925）、RF（AUC=0.897和0.916）和LR（AUC=0.894和0.917）模型预测前列腺癌和临床显著前列腺癌的诊断效能均优于PSA检测（AUC=0.805和0.792）和多参数磁共振检查（AUC=0.773和0.807） （P均<0.05）。SVM模型预测PCa和CSPCa的诊断效能均高于其他模型，但各模型间的诊断效能差异不显著（P均>0.05）。基于验证数据集，SVM的PCa预测模型能够使44例（44/208，21%）患者直接进行前列腺根治术，包括41例（41/65，63%）CSPCa和3例（3/12，25%）GS=3+3的PCa病例，而不过度治疗前列腺穿刺结果为良性的病例，优于ANN（33/208，16%）、RF（17%，36/208）和LR（35/208，17%）模型。 结论机器学习模型预测PCa和CSPCa的诊断效能优于多参数磁共振等单一临床检查，能够使超过60%的CSPCa病例直接进行前列腺根治术，而不会导致过度治疗良性前列腺病例。但需开展前瞻性研究进一步验证该研究结果。

The feasibility of radical prostatectomy without prostate biopsy based on machine-learning models

ObjectiveTo develop machine-learning models for prediction of prostate cancer (PCa) and clinically significant prostate cancer (CSPCa), and explore the feasibility of radical prostatectomy without prostate biopsy based on machine-learning models. MethodsThe total of 688 patients who underwent prostate biopsy according to inclusion criteria between April 2016 and March 2020 were enrolled retrospectively, and the clinical data including age, PSA test, prostate volume (PV), reports of mpMRI examination, and pathological results of prostate biopsy were collected. The diagnostic models of PCa and CSPCa were constructed based on Artificial Neural Network (ANN), Support Vector Machine (SVC), Random Forest (RF) and Logistic Regression (LR) algorithms; and examined with receiver operating characteristic curves. Additionally, we analyzed the feasibility of radical prostatectomy without prostate biopsy based on machine-learning models. ResultsThe ANN (AUC=0.89 and 0.911), SVM (AUC=0.903 and 0.925), RF (AUC=0.897 and 0.916), and LR (AUC=0.894 and 0.917) models outperformed PSA test (AUC=0.805 and 0.792) and mpMRI examination (AUC=0.773 and 0.807) in prediction of PCa and CSPCa (all P<0.05). The SVM models had higher diagnostic accuracy than other models for PCa and CSPCa, but the difference was not significant among the models (each P>0.05). The SVM model for PCa enabled 44(44/208, 21%) patients to undergo radical prostatectomy without prostate biopsy, including 41(41/65, 63%) cases of CSPCa and 3(3/12, 25%) cases of PCa (GS=3+3), without overtreatment of cases with benign biopsy in the validation cohort. ConclusionsThe machine-learning models outperformed single clinical examination such as mpMRI in prediction of PCa and CSPCa, enabling more than 60% of CSPCa cases to undergo radical prostatectomy without prostate biopsy, and would not lead to overtreatment of benign prostate cases. Further prospective validation is required.