Table_1_Machine learning model for the prediction of prostate cancer in patients with low prostate-specific antigen levels: A multicenter retrospective analysis.xlsx

ObjectiveThe aim of this study was to develop a predictive model to improve the accuracy of prostate cancer (PCa) detection in patients with prostate specific antigen (PSA) levels ≤20 ng/mL at the initial puncture biopsy.
MethodsA total of 146 patients (46 with Pca, 31.5%) with PSA ≤20 ng/mL who had undergone transrectal ultrasound-guided 12+X prostate puncture biopsy with clear pathological results at the First Affiliated Hospital of Guangxi Medical University (November 2015 to December 2021) were retrospectively evaluated. The validation group was 116 patients drawn from Changhai Hospital(52 with Pca, 44.8%). Age, body mass index (BMI), serum PSA, PSA-derived indices, several peripheral blood biomarkers, and ultrasound findings were considered as predictive factors and were analyzed by logistic regression. Significant predictors (P < 0.05) were included in five machine learning algorithm models. The performance of the models was evaluated by receiver operating characteristic curves. Decision curve analysis (DCA) was performed to estimate the clinical utility of the models. Ten-fold cross-validation was applied in the training process.
ResultsProstate-specific antigen density, alanine transaminase-to-aspartate transaminase ratio, BMI, and urine red blood cell levels were identified as independent predictors for the differential diagnosis of PCa according to multivariate logistic regression analysis. The RandomForest model exhibited the best predictive performance and had the highest net benefit when compared with the other algorithms, with an area under the curve of 0.871. In addition, DCA had the highest net benefit across the whole range of cut-off points examined.
ConclusionThe RandomForest-based model generated showed good prediction ability for the risk of PCa. Thus, this model could help urologists in the treatment decision-making process.

# 研究目的
本研究旨在构建预测模型，以提升首次前列腺穿刺活检人群中前列腺特异性抗原（PSA，prostate specific antigen）≤20 ng/mL患者的前列腺癌（PCa，prostate cancer）检出准确率。

# 研究方法
本研究回顾性评估了广西医科大学第一附属医院2015年11月至2021年12月期间收治的146例患者，所有患者均接受经直肠超声引导12+X针前列腺穿刺活检且病理结果明确，其前列腺特异性抗原水平≤20 ng/mL，其中前列腺癌患者46例，占比31.5%。验证队列纳入116例来自长海医院的同类患者，其中前列腺癌患者52例，占比44.8%。
本研究将患者年龄、体质量指数（BMI，body mass index）、血清PSA、PSA衍生指标、多项外周血生物标志物及超声影像特征作为预测因子，采用logistic回归进行分析。将P<0.05的显著预测因子纳入5种机器学习算法模型的构建流程。采用受试者工作特征（ROC，receiver operating characteristic）曲线评估模型性能，通过决策曲线分析（DCA，decision curve analysis）估算模型的临床实用性。模型训练过程中采用十折交叉验证法。

# 研究结果
经多因素logistic回归分析，前列腺特异性抗原密度、丙氨酸氨基转移酶与天冬氨酸氨基转移酶比值、体质量指数及尿红细胞水平被确定为前列腺癌鉴别诊断的独立预测因子。相较于其余算法，随机森林（RandomForest）模型展现出最优的预测性能与最高净获益，其曲线下面积（AUC，area under the curve）为0.871。此外，在所有检测的截断值范围内，决策曲线分析的净获益均为最高。

# 研究结论
本研究构建的基于随机森林的模型对前列腺癌风险具备良好的预测能力，因此该模型可辅助泌尿外科医师开展治疗决策工作。