初诊前列腺癌骨转移判别分析模型的建立和验证

发布时间：2018-08-16 19:05

【摘要】：第一部分初诊前列腺癌骨转移判别分析模型的建立和交叉验证目的：对于前列腺癌患者,早期诊断骨转移至关重要。本部分寻找与骨转移相关的临床指标变量,并探索变量的最佳分段方式,从而建立初诊前列腺癌骨转移判别分析模型。方法：本研究纳入2009年1月至2011年11月间在复旦大学附属中山医院泌尿外科初次诊断为前列腺癌的患者,共488例。所有患者在确诊之初都进行了锝99标记的双磷酸盐骨扫描检查。如果骨扫描的结果为疑似骨转移,利用SPECT-CT、CT或MRI对可疑病灶的性质加以进一步确认。每一位患者都有初次诊断时的PSA值、临床TNM分期和前列腺穿刺活检病理Gleason评分。采用多因素回归分析研究上述各临床指标与前列腺癌骨扫描的相关性。然后根据多因素回归结果绘制ROC并计算出AUC。通过判别分析推断需要进行骨扫描的前列腺患者的临床特征。最后对判别分析结果进行leave-one-out法交叉验证。结果：全部488例患者中65例(13.3%)发现有骨转移。cT4期、Gleason评分≥4+3、PSA值是骨转移的独立预测指标。多因素回归模型ROC曲线下面积0.87(95%可信区间为0.83-0.94)。以灵敏度和特异度分别为87.7%和73.1%的切点进行判别分析发现,对于cT1-T3期的患者且Gleason评分≤3+4的患者,若PSA≤132.1 ng/ml骨转移发生率为2%；若PSA132.1 ng/ml,骨转移发生率为25%。对于cT1-T3期的患者且Gleason评分≥4+3的患者,若PSA≤44.5 ng/ml,骨转移发生率为4%；若PSA44.5 ng/ml,骨转移发生率为29%。cT4的患者骨转移发生率为73%。交叉验证结果,灵敏度和特异度分别为86.2%和71.9%,证明模型稳定可靠。结论：cT4期、Gleaso n评分≥4+3、PSA值是骨转移的独立预测指标。本研究构建的判别分析模型能判断初诊前列腺癌患者骨转移的发生率。第二部分潜在预测指标的探索与判别分析模型的优化目的：探索可以优化初诊前列腺癌骨转移判别分析模型的临床指标,并通过优化后的骨转移判别分析模型探讨对于哪些患者初诊时可以不行骨扫描检查。方法：本研究纳入了2009年1月至2012年1月间在复旦大学附属中山医院泌尿外科初次诊断为前列腺癌的患者,共501例。所有患者的病理学诊断为前列腺腺癌。患者的资料包括前列腺穿刺活检病理Gleason评分、临床T分期(根据2002年TNM分期系统)、确诊时血清PSA、ALP、LDH等临床资料被检索收集。其中69例患者在初次诊断时行血清钙、磷检查。所有患者在确诊之初都进行了锝99标记的双磷酸盐骨扫描检查。如果骨扫描的结果为疑似骨转移,利用SPECT-CT、CT或MRI对可疑病灶的性质加以进一步确认。利用多因素回归模型绘制ROC并计算AUC。将ln(PSA+1)、Gleason评分、临床T分期构建多因素回归模型,然后在此回归模型中加入ALP变量得到新的回归模型,新回归模型与原回归模型分别绘制ROC,比较两者的AUC。通过AUC搜索最佳切割值,将PSA和ALP以最佳切割值分为两段,作为判别分析的切点。将判别分析的结果进行表格化。结果：加入ALP变量后ROC曲线下面积(AUC)为0.92,95%可信区间为0.89-0.96,显著大于未加入ALP变量时的曲线下面积(p=0.0012)。对于cTl-T3期的前列腺癌患者如果PSA≤39ng/ml且ALP≤88IU/L,那么骨转移的风险较小,在初诊时不做骨扫描是安全的。其他患者尤其是T4期的患者在初诊时必须做骨扫描。结论：在第一部分筛查模型中加入ALP可以优化此模型,可以提高预测准确性。此模型由单中心的数据建立,需要多中心的研究来验证它的外部真实性。第三部分判别分析模型的外部验证目的：应用外部验证对第二部分中建立的判别分析模型进行评估。方法：外部验证的对象是于2005年3月至2011年3月期间在复旦大学附属肿瘤医院泌尿外科住院治疗的前列腺癌患者,共501例。患者的临床资料包括年龄、Gleason评分、临床分期、PSA值、ALP值等。影像学上诊断骨转移的方法和标准与本研究一致。以第二部分建立判别分析模型作为筛查工具,预测外部数据库中患者在前列腺癌初诊时的骨转移情况。以判别分析结果与实际骨转移情况构建ROC曲线,计算曲线下面积。结果：外部验证的灵敏度和特异度分别达到85.5%和64.0%。ROC曲线的AUC为0.846(95%CI：0.805-0.887),优于其他预测模型。结论：本文建立的判别分析模型稳定可靠,可以应用于临床工作。
[Abstract]:The first part is the establishment and cross-validation of the discriminant analysis model of bone metastasis in newly diagnosed prostate cancer. METHODS: This study included 488 patients with prostate cancer who were initially diagnosed in the Department of Urology, Zhongshan Hospital, Fudan University, from January 2009 to November 2011. All patients underwent bone scan with 99Tc labeled diphosphate at the beginning of diagnosis. Each patient had PSA values at the time of initial diagnosis, clinical TNM staging and Gleason score of prostate biopsy pathology. Correlation between the above clinical parameters and bone scan of prostate cancer was analyzed by multivariate regression analysis. ROC was plotted according to multivariate regression results and AUC was calculated. Results: Bone metastasis was found in 65 of 488 patients (13.3%). CT4, Gleason score (> 4 + 3), PSA value was an independent predictor of bone metastasis. The area was 0.87 (95% confidence interval 0.83-0.94). The sensitivity and specificity were 87.7% and 73.1% respectively. For patients with cT1-T3 and Gleason score < 3+4, the incidence of bone metastasis was 2% if PSA < 132.1 ng/ml, and 25% if PSA132.1 ng/ml. For patients with cT1-T3 and Gleason score < 3+4, the incidence of bone metastasis was 25%. The incidence of bone metastasis was 4% in patients with PSA < 44.5 ng/ml, 29% in patients with PSA < 44.5 ng/ml, and 73% in patients with cT4. The sensitivity and specificity were 86.2% and 71.9% respectively. The model was stable and reliable. Predictive indicators. The discriminant analysis model constructed in this study can determine the incidence of bone metastasis in newly diagnosed prostate cancer patients. Part II Exploration of potential predictors and optimization of discriminant analysis model: To explore the clinical indicators that can optimize the discriminant analysis model of bone metastasis in newly diagnosed prostate cancer, and to discriminate the bone metastasis after optimization. Methods: This study included 501 patients with prostate cancer who were initially diagnosed as prostate cancer in the Department of Urology, Sun Yat-sen Hospital, Fudan University, from January 2009 to January 2012. All patients were pathologically diagnosed as prostate cancer. Gleason score of biopsy pathology, clinical T stage (according to TNM staging system 2002), serum PSA, ALP, LDH and other clinical data were collected at the time of diagnosis. 69 of them were examined for serum calcium and phosphorus at the time of initial diagnosis. All patients underwent technetium-99 labeled bone scan at the beginning of diagnosis. If the bone scan results were Suspected bone metastasis was further confirmed by SPECT-CT, CT or MRI. ROC was plotted by multivariate regression model and AUC was calculated. ln (PSA+1), Gleason score and clinical T-stage were used to construct multivariate regression model. A new regression model was obtained by adding ALP variables into the regression model. The results of discriminant analysis were tabulated. Results: After adding ALP variable, the area under ROC curve (AUC) was 0.92, 95% confidence interval was 0.89-0.96, which was significantly larger than that without ALP variable. If PSA < 39 ng / ml and ALP < 88 IU / L, the risk of bone metastasis is small, and it is safe not to do bone scan at the initial diagnosis. Other patients, especially those with T4, must do bone scan at the initial diagnosis. Conclusion: ALP can be optimized in the first part of the screening model. This model can improve the accuracy of prediction. This model is built from single-center data and needs multi-center research to verify its external authenticity. Part III: The external verification purpose of discriminant analysis model: Applying external verification to evaluate the discriminant analysis model established in Part II. Methods: The external verification object is 200. From March 2005 to March 2011, 501 patients with prostate cancer were hospitalized in the Department of Urology, Tumor Hospital Affiliated to Fudan University. The clinical data included age, Gleason score, clinical stage, PSA value and ALP value. As a screening tool, the ROC curves were constructed to predict the bone metastasis of the patients in the external database at the initial diagnosis of prostate cancer.Results: The sensitivity and specificity of the external verification were 85.5% and 64.0% respectively.The AUC of the ROC curves was 0.846 (95% CI: 0.805-0.887), which was superior to that of the actual bone metastasis. Conclusion: The discriminant analysis model established in this paper is stable and reliable and can be used in clinical work.
【学位授予单位】：复旦大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：R737.25

【参考文献】