- 學位論文
在貝氏統計方法下且考量疾病盛行率的診斷工具網絡統合分析
Diagnostic Test Performance Network Meta-Analysis in Bayesian Approach and Prevalence
摘要
背景 有別於過往介入方式的統合分析，診斷工具的統合分析因為需要同時考慮敏感度與特異度兩個變項，但因為在不同研究之間此二變項並不獨立，因此除了同時考量二者外，估計上亦須考量其相關性。目前常見的診斷工具統合分析的方法為二元模型，而若要進行診斷工具的網絡統合分析，則以ANOVA模型為主。此外因診斷工具的網絡統合分析需要估計許多參數，因此貝氏統計透過資料更新先驗分配，以進行參數估計的方式也較為常用的方法。 目標 在本研究中，將採取貝氏統計的做法，進行診斷工具的網絡統合分析，並以一般診斷工具統合分析的二元模型與網絡統合分析的ANOVA模型為主，且同時將疾病盛行率透過潛在類別分析的方式納入參數考量，以使其能夠更為準確地同時比較多種不同的診斷工具，解決黃金標準可能並不完美的問題。並以SUCRA值、Superiority Index與Youden Index三項指標，評估不同診斷工具之優劣排名。 方法 我們以ANOVA模型為基礎，並在參數中考量疾病盛行率，使模型之結果更加穩定精確。資料驗證上我們以Hoyer and Kuss於2018年研究中比較第二型糖尿病的兩種不同診斷工具之資料，以及Veroniki et al.於2021年比較CIN2+型的子宮頸癌的三種不同診斷工具的研究作為驗證的資料，以確認模型所估計之參數結果之正確性，並嘗試透過不同指標將不同診斷工具之效果進行優劣排名。 結果 本研究中所提出之模型，在Hoyer and Kuss於2018年研究與Veroniki et al.於2021年的研究中皆可求得與論文所發表之研究數據相近之結果。此外對於診斷工具的排名，可以發現SUCRA值因將敏感度與特異度分開考量，可能會導致不同的解讀結果；而Superiority Index則是將敏感度與特異度同時考量，以求出一個理想的排名結果，但沒有絕對範圍；Youden Index則可同時考量敏感度與特異度，並將指標之數值限制於0至1之間，提供一個絕對範圍。 結論 二元模型與ANOVA模型為診斷工具的統合分析提供了優良的架構，以評估多種不同的診斷工具，而將疾病盛行率透過潛在類別分析的方式納入模型參數中，可以為黃金標準並非完全準確時，提供一個解決此問題的方法，在統計估計上可能出現偏差的情況。搭配上貝氏統計的方式，亦能協助我們更有效率地估算參數，以比較不同診斷工具的效果，也提供一個更有彈性的方式，讓使用者可自行調整對參數的分配假設。
並列摘要
Background Unlike the intervention meta-analysis, we need to consider sensitivity and specificity at the same time when we do the diagnostic test performance meta-analysis. However, we not only need to estimate both sensitivity and specificity but also need to estimate their correlation, because they may not be independent across different studies. The general method for diagnostic test performance meta-analysis is the bivariate model. If we want to undertake a diagnostic test performance network meta-analysis to compare several diagnostic tests, the ANOVA model is common method. Besides, the Bayesian approach is also often used for diagnostic test performance network meta-analysis because we can update the prior distribution by the data and estimate the large number of parameters in the network meta-analysis. Objectives We use the Bayesian approach to undertake a diagnostic test performance network meta-analysis in this study. We use the bivariate model and ANOVA model as the basement, and we consider the prevalence by latent class analysis as a parameter in this method to compare several diagnostic tests at the same time to deal with the problem of the imperfect gold standard. We also use three different indexes such as SUCRA, superiority index, and Youden index to rank the different diagnostic tests. Methods We use the bivariate model and ANOVA model as the basis, and add the prevalence as a parameter into these models to obtain more precise estimates. For the demonstration of our models, we use two datasets: one, reported by Hoyer and Kuss, compared two different diagnostic tests for the diagnosis of type 2 diabetes mellitus; the other, reported by Veroniki et al. in 2021, compared three different diagnostic tests for the diagnosis of invasive cervical cancer (CIN2+). We also use different indexes to rank the different diagnostic tests in these researches. Results The results obtained by our model were similar to those given by Hoyer and Kuss or Veroniki et al. In terms of the ranking of diagnostic tests, SUCRA may lead to different decisions because it considers sensitivity and specificity separately. The superiority index considers both sensitivity and specificity together to calculate the ranking, but the range of its value has no limit. Youden index not only considers both sensitivity and specificity simultaneously but also ranges from 0 to 1. Conclusions Both bivariate model and ANOVA model provides a useful statistical framework for assessing the performance of several diagnostic tests. Including the prevalence by latent class analysis as a parameter in the model may provide a solution to the problem of an imperfect gold standard. The Bayesian approach help we estimate the parameters more efficiently and flexibly, and it also allows the users to adjust the prior distributions of model parameters.