RE: A model too far.
RE: A model too far.
复制标题
RE：模型太过分了。
DOI：
10.1093/jnci/dju058
复制
发表时间：
2014
期刊：
影响因子：
--
通讯作者：
Gulati,Roman
中科院分区：
文献类型：
--
作者：
Etzioni,Ruth;Gulati,Roman
Freidlin和Korn关于我们关于前列腺癌过度诊断的文章的社论（1）质疑了我们个性化过度诊断估计在指导治疗决策方面的效用。他们的评论表明，了解过度诊断的风险无助于了解诊断后各种治疗方案的结果。然而，过度诊断告知一个关键的治疗选择的结果-没有治疗。考虑一下一个在筛查测试后被诊断出患有前列腺癌的病人所面临的困境。知道治疗可以降低死于癌症的风险，但治疗可能会让他阳痿或失禁，他应该接受治疗还是不接受治疗？我们的研究结果的价值在于，它们告诉病人如果不治疗他的癌症可能会发生什么。因此，对于活检Gleason评分为7且前列腺特异性抗原为4.5 ng/mL的50岁男性，存在高风险（10分之9的风险），如果他不治疗他的癌症，他将不得不在他生命中的某个时候处理有症状的肿瘤，而对于Gleason评分为6和前列腺特异性抗原为5 ng/mL的80岁男性，这种风险要低得多，约为五分之一。对于这个老年患者，如果他在这一点上什么都不做，他的癌症有80%的可能性永远不会引起症状或问题。这类信息可能非常有助于说服许多年龄较大的低风险患者考虑主动监测，这可能是临床实践中减少过度治疗问题的最重要的变化。Freidlin和Korn还批评了我们的建模方法，声明说，“估计过度诊断的最可靠和透明的方法......是对随机筛选试验的数据进行适当的分析”（1）。然而，他们没有告诉我们什么是“适当的分析”。传统的统计分析在试图从试验中推断过度诊断频率时遇到了问题，因为一个病例是否被过度诊断是不可观察的。曾有人试图使用在筛查组中观察到的过度发生率作为过度诊断的代表，但这种方法通常会产生夸大的估计值（2），特别是在随访不足的情况下，这是所有已发表的前列腺筛查试验的局限性。建模方法深入到经验数据的表面之下，可以应用于试验以及人群发病率数据。这种方法在生物统计学文献中有很长的历史[例如，（3，4）]。总的来说，我们支持对模型持健康的、建设性的怀疑态度。然而，在这种情况下，只有一个模型可以提供对筛查检测到的前列腺癌被过度诊断的可能性的个性化估计，并且所得列线图实质上增加了对前列腺癌的诊断。
Freidlin and Korn’s editorial (1) concerning our article on prostate cancer overdiagnosis questioned the utility of our personalized overdiagnosis estimates for informing treatment decision making. Their comments suggest that knowing the risk of overdiagnosis does not help to inform about the outcomes of various treatment options after diagnosis. However, overdiagnosis informs about the outcome of a key treatment option—that of no treatment. Consider the dilemma faced by a patient who has just been diagnosed with prostate cancer after a screening test. Knowing that treatment can reduce the risk of dying from cancer but the treatment could leave him impotent or incontinent, should he be treated or not? The value of our results is that they inform the patient about what might happen if his cancer were to be left untreated. Thus, for a 50-year-old man with a biopsy Gleason score of 7 and a prostate-specific antigen of 4.5 ng/mL, there is a high risk (risk of 9 of 10) that he will have to deal with a symptomatic tumor at some point in his life if he does not treat his cancer, whereas for an 80-year-old man with a Gleason score 6 and a prostate-specific antigen of 5 ng/mL, this risk is much lower, approximately one in five. For this older patient, if he does absolutely nothing at this point, there is an 80% chance that his cancer will never cause symptoms or problems. This kind of information could be very helpful in persuading many older, low-risk patients to consider active surveillance, which is likely to be the single most important change in clinical practice to reduce the problem of overtreatment.Freidlin and Korn also critique our modeling approach, stating that “the most reliable and transparent approach to estimating overdiagnosis… is with an appropriate analysis of data from a randomized screening trial”(1). However they do not tell us what an “appropriate analysis” might be. Conventional statistical analysis runs into problems when attempting to infer overdiagnosis frequencies from trials because whether a case has been overdiagnosed is not observable. There have been attempts to use observed excess incidence in the screened arm as a proxy for overdiagnosis, but this approach, because it typically produces inflated estimates (2), particularly under insufficient follow-up, which is a limitation of all published prostate screening trials. A modeling approach goes beneath the surface of empirical data and can be applied to trials as well as to population incidence data. This approach has a long history in the biostatistics literature [eg,(3, 4)]. In general, we support a healthy and constructive skepticism of models. Yet, in this this situation, only a model can provide personalized estimates of the chance that a screen-detected prostate cancer has been overdiagnosed, and the resulting nomogram adds materially to the