從台灣健保資料庫找尋導致副作用的藥物交互作用配對

Abstract

藥物在上市前都會經過臨床試驗的階段，但受制於試驗時間長度及受試者人數的關係，依舊難以完整得知藥物在人體身上的作用效果，因此，藥物的上市後監督就變得非常重要。藥物主動監視（Pharmacovigilance）是指對上市藥品的實際使用狀態進行分析、評估，並進一步預防可能發生的副作用。有多種原因能導致副作用的發生，其中一項即為藥物交互作用。因此若能偵測藥物交互作用的行為，即能避免損害發生。藥物主動監視主要可透過兩種資料來源進行分析，分別是不良反應通報系統，和電子健康紀錄。雖然不良反應通報系統已被研究證明為一可信賴的資料來源，但其存在一些問題，例如通報率低、涵蓋人口不明等，電子健康紀錄逐漸成為替代的分析資料。 在本篇研究中，使用了台灣健保資料庫來分析藥物交互作用的行為，希望能預測出那些導致副作用的藥物交互作用配對，稱之為藥物交互作用訊號。我們提出了一個方法，藉由排序學習法結合多個常用的偵測指標，去預測藥物交互作用訊號的名單，並對其依可疑程度進行排序。由於此方法為一監督式學習法，我們邀請了藥學專家幫忙進行資料標註。本研究選擇心血管疾病和肝毒作為檢測的副作用疾病。研究結果顯示，我們所提出的方法能夠有效的提升藥物交互作用偵測的準確度，並能將此預測的排序名單，提供給醫藥專家做進一步的驗證及分析。

It is difficult to know the complete safety information about a new drug during pre-marketing trails. Because of this, post-marketing monitoring of drug safety becomes extremely important. Pharmacovigilance (PhV) is the pharmacological science relating to the collection, detection, assessment, monitoring, and prevention of adverse effects with pharmaceutical products. One of the causes of adverse events is drug-drug interactions (DDIs). If we can detect the behaviors of DDIs, the damage of adverse events can be decreased. Two major data sources that can be employed for PhV include spontaneous reporting system (SRS) and electronic health record (EHR). Since SRS has some inherent problems such as underreporting or unknown denominator, recent research attention has started to focus on the use of EHR for PhV purposes. In our study, the objective is to identify DDIs that cause adverse events, which we call as DDI signals. We propose a novel method to predict and rank signals by using Taiwan’s National Health Insurance Research Database. We employ the learning to rank approach and exploit multiple measures as the predictors for ranking signals. For training and evaluation purposes, we invite domain experts to annotate datasets for ranking-model training and evaluation purposes. Cardiovascular events and hepatotoxicity are selected to be the detected diseases. The empirical evaluation results show that our proposed method has the ability of predicting and ranking DDI signals, which can support pharmaceutical professionals to do further analysis and investigation.