使用序貫網絡統合分析比較牙周再生手術的成效

Abstract

背景 統合分析（meta analysis）在近四十年被普遍使用在統合隨機分派臨床試驗（randomize controlled trial [RCT]）的證據。而為了確認是否有必要針對某一項特定主題重複不斷地做研究，有學者提出了累積統合分析，藉由一系列依照時序進行的統合分析，加上森林圖的呈現，讓研究者可能在某一個時間點，看到兩種干預效果有明顯的差異。但這樣一系列的檢定，可能讓整體型一錯誤發生的機率上升，而檢定力也可能無法達到我們研究開始前所設定的標準。因此有學者將臨床試驗中常使用的期中分析（interim analysis）的想法引入統合分析，提出了試驗序貫分析方法（trial sequential analysis）以及近一步將間接證據考慮進去的序貫網絡統合分析（sequential network meta analysis），來校正統合分析中整體可容忍的型一錯誤。

目標 本研究利用序貫網絡統合分析，分析牙釉基質衍生物（enamel matrix derivatives）及引導組織再生術（guided tissue regeneration）的相關治療對於牙周病的改善是否有顯著差異，並且和一般的網絡統合分析結果進行比較。

材料及方法 本研究先對於統合分析、網絡統合分析及累積統合分析做介紹，再回顧期中分析中所使用的分組序貫方法（group sequential methods），例如α花費函數（α-spending function）及三角測試（triangular test），最後則介紹試驗序貫分析及序貫網絡統合分析。此外，本研究也針對序貫網絡統合分析所使用的R套件（sequentialnma）進行些微的修改，並用探討模擬資料使用序貫網絡統合分析的表現。而資料分析部分，我們使用Tu等人2012年在Journal of Clinical Periodontology所發表之研究的資料(Tu et al., 2012)，並搜尋2011年1月到2017年12月的相關文獻進行資料的更新。將更新後的資料使用序貫網絡統合分析，分析不同的治療在改善牙周囊袋深度（probing pocket depth [PPD]）及臨床附著水平（clinical attachment level [CAL]）上，是否有顯著差異。

結果 我們總共納入了58篇研究進行分析。而序貫網絡統合分析和一般網絡統合分析的結果差異不大。在改善PPD及CAL上，我們得到除了和使用牙釉基質衍生物加上骨移植體並使用富血小板血漿治療（enamel matrix derivatives with platelet rich plasma and bone graft；EMD+PRP+BG）沒有顯著差異之外，其他六種治療都明顯優於翻瓣手術（flap operation）治療（雖然使用EMD+PRP+BG治療比翻瓣手術能多減少0.95公厘的牙周囊袋深度，但其信賴區間及重複信賴區間分別為[-0.88, 2.78]及[-2.31, 4.21]，顯示兩組治療沒有顯著差異）。然而在其他的兩兩成對比較，都沒有任何治療顯著的優於另一治療。不過，在效應值為PPD改變量時，用牙釉基質衍生物（EMD）治療和使用引導組織再生術（GTR）治療的比較上，使用序貫網絡統合分析會得到和一般網絡統合分析下不同的結論：如果在一般網絡統合分析下，最後會得到GTR 治療優於EMD 治療的結論(其信賴區間為[0.06, 0.86])，但在序貫網絡統合分析之下，這兩種治療就沒有顯著差異了(其重複信賴區間為[-0.02, 0.94]).，並且需要進行後續的研究。至於在資訊的累積上，有統計上顯著差異的治療比較，相對於不顯著的治療比較都累積較高的資訊量（從0.77到2.52這個範圍中）。

結論 序貫網絡統合分析和一般網絡統合分析的結果雖然差異不大，但使用序貫網絡統合分析能讓我們在分析時控制整體研究的型一錯誤，使我們不會過早下錯誤的結論。並且透過證據的累積，我們能知道目前我們所有的資訊量，以及後續是否需要再進行相關研究，以節省研究的資源。

Background: Meta-analyses and network meta-analyses (NMA) have become popular for scholars to evidence synthesis to attain a robust conclusion. To understand the trend in the results, cumulative meta-analysis has been proposed. However, cumulative meta-analysis undertakes a series of statistical tests. While the type I error rate is usually set at 5%, multiple testing gives rise to inflated type I error rate.

Objective: This thesis aims to compare the effects of different periodontal regenerative therapies on the treatment of teeth with periodontal diseases by using sequential network meta-analysis (SNMA), and also compares the SNMA results with the general network meta-analysis’s.

Materials & Methods: To deal with the inflation of type I error rate, the group sequential methods (e.g. the α-spending function, the Whitehead triangular test), which were used in the interim analysis in the clinical trials, have been introduced into meta-analyses. Trial sequential analysis (TSA) proposed by Wetterslev and his colleagues, and sequential network meta-analysis proposed by Nikolakopoulou and her colleagues, are two useful methods to resolve this problem. SNMA also takes indirect evidence into account. In the thesis, we modified the “sequentialnma” package in R which was then used to analyze updated data from a network meta-analysis on the effects of different periodontal regenerative therapies in 2012 (Tu et al., 2012), by searching new studies from January, 2011 to December, 2017. And then we use SNMA, TSA and triangular test to compare guided tissue regeneration (GTR), enamel matrix derivatives (EMD) and their combination therapies regarding the probing pocket depth (PPD) reduction and clinical attachment level (CAL) gain.

Results: A total of 58 studies were included in this thesis, and SNMA and NMA yield similar results. In terms of PPD reduction and CAL gain, the flap operation (FO) was significantly less effective than other treatments except for the EMD with platelet-rich plasma and bone graft therapy. Although the EMD with platelet-rich plasma and bone graft therapy had on average attained 0.95 mm greater PPD reduction than FO, but the confidence interval [-0.88, 2.78] and the repeated confidence interval [-2.31, 4.21] showed no statistical significance in their difference. No significant difference was found in other pairwise comparison. SNMA showed that the difference in PPD reduction between EMD and GTR was not statistically significant (the repeated confidence interval was [-0.02, 0.94]), while NMA showed that GTR attained greater PPD reduction than EMD (the confidence interval was [0.06, 0.86]). As for the accumulation of information in the analysis, the statistically significant comparisons usually had a greater proportion of the accumulated information (ranging from 0.77 to 2.52) than the nonsignificant one.

Conclusion: Although SNMA and NMA achieved similar results, using SNMA could help control the type I error rate thereby reducing the risk of false-positive findings. SNMA also summarizes information accrued from the beginning up to a particular time, and decision can then be made with regards to whether or not evidence is sufficient to conclude the superiority of one treatment over another or the equivalence of two treatments. Resources can therefore be directed to where it is most needed in the future.