在元件網絡統合分析中重建斷裂網絡之條件及證據結構視覺化之方法

Abstract

網絡統合分析是用於比較多種治療相對效果的工具。最常使用的網絡統合分析模型（Lu & Ades模型）無法用於分析網絡斷裂的資料。儘管陸續有不同的解決方式被提出，但他們都有各自的缺點。而元件網絡統合分析便是其中一種潛在的分析斷裂網絡的方式。 元件網絡統合分析是基於相加性假設來分析多種組合式治療的元件效果。近年有文獻指出，元件網絡統合分析有可能使斷裂的網絡重新建立起連結，其必要條件是斷裂網絡中的子網絡之間要有共同元件；然而這並非唯一的必要條件，即使滿足前述的條件也不能保證網絡一定能透過元件網絡統合分析重建。因此，本研究的第一個目標是找齊所有的必要條件。我們證明要反應證據結構的M矩陣的秩等於元件數才可使所有的元件皆可估計。當斷裂網絡滿足這兩個條件，即可使用元件網絡分析進行重建。 接著，我們藉由發展一種視覺化證據結構的繪圖方式，來研究如何檢查連結不同子網路欠缺何種證據。然而標準的網絡圖採用治療當作結點，無法很好的呈現出元件網絡統合分析的證據結構，尤其是資料一複雜就難以判讀。因此，本研究的第二個目標是使用線性信號流圖以元件為結點呈現它們彼此的關係。新的繪圖方式能清楚地呈現元件本身的證據分布情況。由於是根據標準網絡圖轉換而來，我們也發現標準網絡圖其實可視作我們新圖的特例。而若資料複雜度較高使圖片變得難以判讀，還可以在繪圖前採高斯消去法簡化圖上的訊息，以輔助使用者解讀證據分布之情況。

Network meta-analysis (NMA) is used to compare the relative effects of multiple treatments. The most popular network meta-analysis model, namely the Lu & Ades model, cannot analyze a disconnected network. Several solutions have been proposed, but they all have their own disadvantages. One potential solution to analyzing a disconnected network is component network meta-analysis (CNMA). CNMA estimates the effect of each component within multicomponent interventions under the assumption of additive component effects. A recent article pointed out that CNMA could reconnect disconnected networks when distinct subnetworks share common components. However, this condition itself is insufficient for connecting broken networks. Our first objective, therefore, is to find the necessary conditions. We proved that the rank of the design matrix M needs to be equal to the number of components for identifying the effects of components. A disconnected network will be reconnected as it these two conditions are satisfied. We, then, investigated how to identify the missing evidence required to connect distinct subnetworks by developing a graphical tool for visualizing the evidence structure. The standard network plot uses treatments as its nodes, which is not suitable for visualizing the evidence structure in CNMA, particularly when the network plot is complicated. Therefore, our second objective is to propose the use of the linear signal-flow graph, which illustrates the relations between components as connections between nodes. Our proposed graph provides clear information about the relations among components within the networks of evidence, and it can be shown that the standard network plot is a special case of our proposed graph. Moreover, we can use gaussian elimination to extract information within our graph to help the user elaborate the evidence structure.