於健保資料庫中偵測藥物不良反應

Abstract

藥物不良反應（Adverse Drug Reactions，簡稱ADRs）係指接受藥物治療後所產生的嚴重健康危害。更由於ADRs是當今主要死因之一，故妥善監視上市後藥物成為一重要課題。然而，傳統的失衡分析法（disproportionality analysis）與貝氏偵測方法（Bayesian signal detection）仰賴預先收集的ADR通報案例，以及需事先定義、無統一標準的門檻值，其偵測結果也經常無法一致。另一方面，用以進行偵測的資料集長久受限於兩個資料庫—美國FDA之FAERS與WHO之VigiBase，於這些資料庫的偵測也存在諸多困難。 為解決上述問題，本研究使用全民健康保險研究資料庫，以一週為單位聚合每位病患之歷史就診紀錄後，建立藥物與診斷先後關係。我們並提出一結合三種偵測分數：回歸t值（REG）、通報相對比例值（PRR）與通報相對勝算比（ROR）作為輸入特徵的新模型，用以偵測藥物不良反應。實驗結果顯示，相較單獨使用一種分數，結合三種偵測分數的新模型之準確度（Accuracy）最高有9.5%的提升。

Adverse Drug Reactions (ADRs) are fatal health problems due to medical treatments. ADRs are leading cause of death, and thus it is crucial to properly monitor post-marketing drugs. However, traditional disproportionality analysis and Bayesian signal detection depend on pre-collected ADR reports and a not universal, predefined threshold; the results are often inconsistent. Moreover, the available data sources were limited to two databases — U.S. FDA’s FAERS and WHO’s VigiBase; there are also several difficulties when detecting ADRs in these databases. To address above problems, in this study, we proposed a model combining three detecting scores: regression’s t-value (REG), proportional reporting ratio (PRR), and reporting odds ratio (ROR), as features for detecting serious drug-ADR pairs from one-week aggregated patient-week information with precedence relationship between drugs and diagnoses, in an health insurance claims database NHIRD (National Health Insurance Research Database). We demonstrated that the proposed combined score led to an improvement (up to 9.5%) of signal detection accuracy over applying each of score independently.