網絡統合分析排名之可信度

Abstract

近年來國際上的臨床指引廣泛採用網絡統合分析的結果去提供疾病治療的建議。透過網絡統合分析，整合多個研究的直接證據與間接證據，去估算多種治療方法或者介入措施之間的差異，藉此可以提供實證醫學一個有利的工具來填補當前的知識缺口。然而，由於網絡統合分析中所包含的多為有效的治療方式，因此它們之間的效果的差異往往很小或是沒有達到統計上顯著，故其結果並不容易解讀。因此，就有研究者提出使用排名的方法來去簡化治療之間比較結果的解讀。 透過排名，可以讓網絡統合分析結果的資訊簡單化，也是將實證數據轉化為臨床實務的一種方式。排名讓複雜的網絡統合分析結果容易解讀，然而，不論資料多寡，只要可以進行網絡統合分析，就能取得其治療的排名。但是，在排名上有差別的兩個治療，並不表示他們之間的差異就很顯著。因此，使用排名卻不報告排名的可信度，往往會導致誇大解讀不同介入或治療之間的差異。 目前排名可信度的評估方法包括不確定性評估和穩健性評估。當前排名不確定性的評估方法，會受到網絡所包含的治療數目所影響，因此被批評此指標是資訊缺乏的。而穩健性評估與不確定性評估之間的關聯為何，目前並未有定論。因此，本論文旨在建立評估網絡統合分析排名可信度的方法，以強化對網絡統合分析的解讀和應用。本論文所探討的問題如下所述： 1. 發展網絡統合分析中治療排名不確定性的替代指標。 2. 探討排名的穩健性和不確定性關聯性。 針對以上研究問題，首先，本論文提出應用標準化熵這個度量，來將每種治療的排名機率分佈轉換為一個數值指標，以促進對治療排名不確定性的精確解讀。標準化熵是一個介於0到1之間的指標，越大表示不確定性越高。與傳統的指標相比，此指標不會受到包含在網絡當中的治療數目多寡影響，因此，它可用於比較網絡統合分析中不同治療，或是和不同網絡統合分析之間治療排名的不確定性。而本論文利用網絡統合分析資料庫，使用標準化熵評估157篇已發表的網絡統合分析，其中排名不確定性高的網絡統合分析占約三分之二。此外，本論文利用已發表的網絡統合分析，來探討排名不確定性和穩健性的關係。從結果看到，與預期相符的是，排名不確定性很低時，相對的排名穩健性也很高。然而，排名穩健性高並非總是對應於低不確定性，具有高穩健治療的排名也可能同時具有高不確定性。因此，當穩健性高的時候，並不表示此排名未來不容易改變，只能說在此網絡所包含的試驗中，沒有單一一個試驗是會對排名有很大影響的。 在報告排名時，利用標準化熵來呈現排名不確定性可以讓我們避免對排名的過度解度。目前已發表的網絡統合分析，排名的不確定性極高，表示其排名可能會在未來有新的試驗加入時改變。而透過一次排除一個試驗來看排名穩健性的方法，只能審視目前包含的試驗是否對排名會有很大的影響，並非與排名不確定性有絕對的相關性。

In recent years, network meta-analysis (NMA) has been widely used to formulate recommendations in the guidelines for managing diseases. NMA combines both direct and indirect evidence to compare multiple treatments and has been shown to be a useful tool for bridging the knowledge gap in evidence-based medicine. Since the differences among active treatments in the efficacy or harm are likely to be small, researchers develop methods to rank treatment for aiding the interpretation of treatment comparisons. Ranking makes information from NMA simpler and is also a way to translate evidence into clinical practice. However, although ranking facilitates the interpretation of complex results from NMAs, its reliability has caused much controversy. As ranking can always be obtained from NMA, the difference in ranking does not mean that difference between treatments is statistically significant. Therefore, using rankings without reporting the reliability of the rankings may either make interpretation difficult or exaggerate the small differences between treatments. Currently, ranking uncertainty and ranking robustness are two methods for evaluating the reliability of ranking. However, the current method used to evaluate the uncertainty of ranking would be affected by the number of treatments included within the network. Therefore, it is criticized as uninformative. The association between ranking uncertainty and ranking robustness has not been fully explored. Thus, the objective of this dissertation is to develop methods to facilitate the interpretation and application of NMA rankings. This dissertation aims to address the following objectives: 1. Develop an alternative method to measure the uncertainty of treatment ranking from NMA. 2. Explore the association between the uncertainty of ranking and robustness of ranking For the first objective, I proposed Normalized Entropy, which transforms the distribution of ranking probabilities into a single quantitative measure to facilitate a refined interpretation of uncertainty of treatment ranking. I showed that as Normalized Entropy ranges from 0 to 1 and is independent of the number of treatments, it can be used to compare the uncertainty of treatment rankings within an NMA and between different NMAs. Normalized Entropy is an alternative tool for measuring the uncertainty of treatment ranking by improving the translation of results from NMAs to clinical practice and avoiding nave interpretation of treatment ranking. I also evaluated the uncertainty of ranking for 157 published NMAs. Among them, two-thirds of NMAs have high or very high ranking uncertainty. In the results of the second objective of the dissertation, the association between uncertainty and robustness of ranking was explored. The results showed that low uncertainty corresponds to high robustness. When the uncertainty of ranking is very low, treatment ranking is unlikely to be altered by deleting a trial from the complete data. However, good robustness of ranking does not always correspond to low uncertainty. NMA with robust treatment ranking may have high uncertainty of treatment ranking. Therefore, if the network does not contain a trial that significantly impacts the ranking, even if the uncertainty is high, the ranking robustness can still be high. The high robustness of ranking does not mean that the ranking will not be easily changed when new trials are added in the future, but it means that the network does not contain trials that have a significant impact on the treatment ranking. When reporting rankings, using Normalized Entropy to present ranking uncertainty prevents us from nave interpretation of treatment ranking. Among the current published NMAs, most of them have high uncertainty of ranking, and their rankings may have a higher possibility to be changed when new trials are added into the network in the future. The robustness of ranking, which is evaluated by the leave-one-trial-out approach to identify trials included in the network that substantially influence the treatment ranking, is not entirely related to the uncertainty of ranking.