David Geffen UCLA School of Medicine, Los Angeles, California, USA.
Olive View UCLA Medical Center, Sylmar, California, USA.
Nutr Clin Pract. 2019 Feb;34(1):60-72. doi: 10.1002/ncp.10227. Epub 2018 Dec 20.
Evidence-based medicine (EBM) has become a fixture in today's medical practice. Evidence consists of memorialized observations and should be contrasted with dogmatic pronouncements and/or hypotheses. Evidence has varying degrees of reliability. The randomized clinical trial (RCT) or a systematic review of RCTs is accorded the highest level of credibility and expert opinion the lowest. This ranking reflects the internal validity (degree to which factors in the study interfere with the gathering or interpretation of the observations) of the study design; more valid designs are more credible. The provision of healthcare requires an almost constant assessment of evidence. In so doing, there are a number of principles of EBM that need to be kept in mind: Association can never prove causation. Various methodologic biases can influence conclusions made in both RCTs and observational studies. The strength of RCTs is in the elimination of confounding bias. Surrogate outcomes must be validated in RCTs assessing how they are changed compared with the clinical outcomes. Subgroup analyses cannot prove hypotheses although they can generate them. P < 0.05 is not the same as truth. Type I errors are more likely to occur when multiple analyses are performed, when trials are prematurely stopped for perceived benefit when there was no a priori plan to do so, or in small papers with dramatic results that are selectively published. The failure to find a difference does not mean that no difference exists (type II error).
循证医学(EBM)已成为当今医学实践的固定内容。证据包括被记录的观察结果,应与教条式声明和/或假设形成对比。证据具有不同程度的可靠性。随机临床试验(RCT)或 RCT 的系统评价被认为是最可信的,而专家意见则被认为是最低级别的。这种排名反映了研究设计的内部有效性(研究中干扰观察结果收集或解释的因素的程度);更有效的设计更可信。医疗保健的提供需要对证据进行几乎持续的评估。在这样做的过程中,需要记住循证医学的一些原则:
关联永远不能证明因果关系。各种方法学偏差会影响 RCT 和观察性研究得出的结论。
RCT 的优势在于消除混杂偏差。
在评估替代终点如何与临床结局相比发生变化的 RCT 中,替代终点必须得到验证。
亚组分析不能证明假设,尽管它们可以产生假设。
P < 0.05 与真相不同。当进行多次分析、试验因预期获益而提前停止而事先没有计划这样做时、或在具有选择性发布的引人注目的结果的小论文中,更容易发生 I 型错误。未发现差异并不意味着不存在差异(II 型错误)。