Poulakis Vassilis, Witzsch Ulrich, de Vries Rachelle, Emmerlich Volker, Meves Michael, Altmannsberger Hans-Michael, Becht Eduard
Department of Urology, Krankenhaus Nordwest, Teaching Hospital of the Johann-Wolfgang-Goethe-University, Frankfurt, Frankfurt/Main, Germany.
Urology. 2004 Dec;64(6):1165-70. doi: 10.1016/j.urology.2004.06.030.
To develop and test an artificial neural network (ANN) for predicting biochemical recurrence based on the combined use of pelvic coil magnetic resonance imaging (pMRI), prostate-specific antigen (PSA) measurement, and biopsy Gleason score, after radical prostatectomy and to investigate whether it is more accurate than logistic regression analysis (LRA) in men with clinically localized prostate cancer.
We evaluated 191 consecutive men who had undergone retropubic radical prostatectomy for clinically localized prostate cancer. None of the men had lymph node metastasis as determined by adequate follow-up and pathologic criteria. The preoperative predictive variables included clinical TNM stage, serum PSA level, biopsy Gleason score, and pMRI findings. The predicted result was biochemical failure (PSA level of 0.1 ng/mL or greater). The patient data were randomly split into four cross-validation sets and used to develop and validate the LRA and ANN models. The predictive ability of the ANN was compared with that of LRA, Han tables, and the Kattan nomogram using area under the receiver operating characteristic curve (AUROC) analysis.
Of the 191 patients, 57 (30%) developed disease progression at a median follow-up of 64 months (mean 61, range 2 to 86). Using all the input variables, the AUROC of the ANN was significantly greater (P <0.05) than the AUROC of LRA, Han tables, or the Kattan nomogram for the prediction of PSA recurrence 5 years after radical prostatectomy (0.897 +/- 0.063 versus 0.785 +/- 0.060, 0.733 +/- 0.061, and 0.737 +/- 0.071, respectively). Removing the pMRI findings from the previous models, the AUROC of the ANN decreased statistically significantly (P <0.05) and was comparable to the AUROC of conventional predictive tools (P >0.05).
Using the pMRI findings, the ANN was superior to LRA, predictive tables, and nomograms to predict biochemical recurrence accurately. Confirmatory studies are warranted.
开发并测试一种人工神经网络（ANN），用于在根治性前列腺切除术后，基于盆腔线圈磁共振成像（pMRI）、前列腺特异性抗原（PSA）测量值和活检Gleason评分的联合使用来预测生化复发，并调查其在临床局限性前列腺癌男性患者中是否比逻辑回归分析（LRA）更准确。
我们评估了191例因临床局限性前列腺癌接受耻骨后根治性前列腺切除术的连续男性患者。根据充分的随访和病理标准，这些患者均无淋巴结转移。术前预测变量包括临床TNM分期、血清PSA水平、活检Gleason评分和pMRI结果。预测结果为生化失败（PSA水平为0.1 ng/mL或更高）。将患者数据随机分为四个交叉验证集，用于开发和验证LRA和ANN模型。使用受试者操作特征曲线下面积（AUROC）分析，将ANN的预测能力与LRA、Han表和Kattan列线图的预测能力进行比较。
在191例患者中，57例（30%）在中位随访64个月（平均61个月，范围2至86个月）时出现疾病进展。使用所有输入变量，对于预测根治性前列腺切除术后5年的PSA复发，ANN的AUROC显著高于LRA、Han表或Kattan列线图（P<0.05）（分别为0.897±0.063、0.785±0.060、0.733±0.061和0.737±0.071）。从前述模型中去除pMRI结果后，ANN的AUROC在统计学上显著下降（P<0.05），且与传统预测工具的AUROC相当（P>0.05）。
使用pMRI结果，ANN在准确预测生化复发方面优于LRA、预测表和列线图。有必要进行验证性研究。
