2018-08-012024-05-17https://scholars.lib.ntu.edu.tw/handle/123456789/683003摘要：新上市的藥物雖然要經過大型臨床試驗證實療效與安全，但是臨床試驗入組病患篩選嚴格，和真實世界實際治療的族群有明顯差別。不同族群藥物療效可能有所差異，舉例而言，新型抗血小板藥物Ticagrelor在歐美族群的第三期試驗顯示較Clopidogrel有較強的抗血小板效果，但是最近以日本為主的東亞跨國臨床試驗卻顯示Ticagrelor並沒有比Clopidogrel有較佳的心血管保護效果。因此，上市後持續應用大型資料庫長期追蹤病患分析相對效用與安全性，提供真實世界證據，對臨床治療可以提供臨床試驗所無法提供的重要醫療決策資訊。傳統藥物相對效用分析以多變數迴歸分析校正干擾因子。然而受限於領域知識和干擾因子資訊的收集不足，殘存干擾的疑慮一直不能完全排除。最近,哈佛大學 Schneeweiss教授使用資料探勘技術自動篩選干擾因子, 突破了傳統方法，可以選取多達500個干擾因子，讓干擾的校正有大幅的進展。然而這個流程只有在數萬筆大小的資料庫驗證過，若需應用於千萬病患的大型資料庫，則會有運算效率的問題 。為了達成實時分析決策的輔助，本計畫第一個目標是利用資料統整(ICD-9 grouping)或維度縮減(dimension reduction)縮短電腦自動篩選干擾因子的分析時間。本計畫要解決的第二個問題是相對效用與相對安全的問題，新藥追求更佳療效的同時也往往會伴隨較高的副作用，效用與安全性的取捨形成了困境。例如更強的抗血小板藥物有更強的保護效果，但是卻也可能造成較多的出血副作用 (如Ticagrelor)。整體利弊取捨形成了臨床上醫療決策的困境。本計畫降嘗試用新的分析方法，整合效用與傷害成一個衡量指標，進行相對效用對傷害分析(Comparative effectiveness-to-harm analysis)或是效益風險分析，新發展的指標以死亡率為基礎，給予效用和傷害不同的權重，再加以整合，因此可以計算出新藥在人群中的整體淨效益。相關概念在2015年才剛發展，還有許多延伸應用與驗證空間，我們將率先應用於分析台灣健保資料庫或是電子病歷，探討不同指標的相關意義。最後, 本計畫的第三目標是整合資料探勘與效益風險分析，建立一個全自動化的分析程式，應用於Ticagrelor與Clopidogrel的監測，進行相關整體人群風險效益評估。<br> Abstract: Although the approval of a new drug requires a phase III trial, expected results from clinical trial are usually not observed in the real world practice. Differences in the patient population between clinical trials and real-world practices are thought to be the major cause. For example, several US-based trials of ticagrelor, a newly launched antiplatelet medication, reported a better preventive capability for major adverse cardiac events than clopidogrel. However, in a recent multi-national observational study in Asia, Ticagrelor showed neither more effective nor safer than clopidogrel. Therefore, the US Food and Drug Administrative authority has embarked on a change on the new drug evaluation process. Unlike the past regulation where pharmaceutical company-led phase III trial is a major determinant for the benefit-risk profile of a new drug, the new paradigm extends the evaluation process to the real world evidence in the post-marking period. As most of the real world evidence is generated by observational studies, the control of confounding become the major issue. Conventionally, confounders are usually defined by context knowledge. Limited by the domain knowledge, the number of confounders are usually limited, leaving residual confounding a serious concern. Real-world database, such as EMR or administrative databases, collected computerized data in an automatic fashion and thus formed a huge database containing far more data than conventional clinical trials or clinical studies. Recently, Schneeweiss et al developed a new automated confounder selection algorithm, i.e. the high dimensional propensity score (hdPS) algorithm. Although the hdPS has realized the standardized analysis in large health database, it faces a problem with algorithm efficiency. The calculation may take up to weeks to months for a large database involving tens of millions of people. Therefore, the first aim is to try to improve the efficiency of hdPS. We will apply the dimension reduction technique such as principal component analysis and ICD-9 codes collapsing, to enhance the computational efficiency. The second aim we want to address is the effectiveness-to-harm problem. A new drug is usually more effective than the standard treatment, but it may also cause some unexpected adverse events. For example, new anti-platelet agents are more potent, but also carry higher risk of significant bleeding events. Traditional pharmacoepidemiology approach compares the effectiveness and safety separately. However, the medical decision for an individual requires consideration of both benefits and harms. A metric that measure effectiveness-to-harm ratio simultaneously is more useful in guiding medical decision. Thus, the second aim of this study is to develop an effectiveness-to-harm metric. We will first determine the impact of stroke/MI or bleeding events on the 1-year survival and the derived hazard ratio will be used as the relative weight for weighted NNT (number needed to treat) to NNH (number needed to harm) ratio metrics. The last aim is to integrate the previous two aims into an automated real time comparative effectiveness-to-harm analysis system and apply it in both EMR and administrative database. We will use the clopidogrel and ticagrelor as a proof-of-concept pilot study to demonstrate the feasibility of automatic effectiveness-to-harm analysis.真實世界證據高維度傾向計分維度縮減效益風險分析相對效用對傷害分析real-world evidencehigh-dimensional propensity scoredimension reductionbenefit risk analysiscomparative effectiveness-to-harm analysis