目的：心臟血管內科門診處方向來有高的潛在性藥物交互作用發生率，一旦發生藥物交互作用，將使預期療效無法達到，甚或引發其他疾病導致死亡。本研究期以決策樹技術發掘影響藥物交互作用有意義的資訊，俾利資訊科技建置有效的檢核提示系統，以提升病人的用藥安全。方法：2004年全民健保研究資料庫門診處方及2004年衛生署藥物交互作用資料庫為研究資料。以Clementine7.2建立決策樹演算法C5.0及CART分類模式。結果：資料庫門診處方以心臟血管內科藥物交互作用的發生率19%為最高，以年齡65~74歲(26.8%)及75歲以上(24.8%)為主要族群。就心臟血管內科交互作用發生率而言，1至5級分別為12.9%、51.4%、4.7%、14.9%及16.1%，可能具致死性或造成永久性傷害及導致病情惡化嚴重度1及2級共占64.3%。藥物交互作用組合中以Digitalis Glycosides與Loop Diuretics的組合發生率28.16%為最多。資料探勘決策樹演算法C5.0的藥物交互作用預測正確率較CART為高，C5.0分類模式的最重要預測變數為主診斷，反應出主診斷病況在藥物交互作用上所扮演的重要角色。建議醫院成立臨床專家小組，建立客製化的藥物交互作用資料庫，以共同促進病人用藥的安全。 OBJECTIVE: Cardiological prescriptions are with high risk of potential drug-drug interactions (DDIs). DDIs will enable pharmacotherapeutic failure, and be associated with morbidity and mortality. The objective of the present study was to build a decision tree to determine the important factors for the effective DDIs screening software development. METHODS: The 2004 claims data of the Bureau of National Health Insurance (BNHI) from hospitals throughout Taiwan which contains ambulatory prescriptions, therapeutic and registration files were the data source. The 2004 DDIs Database developed by the Department of Health was used to select potentially harmful drug combinations in the outpatient setting. Clementine 7.2 was used for the classification task of data mining in order to determine DDIs predictors for drug safety. RESULTS: In ambulatory prescriptions, 19 % (2,307/12,350) cardiological prescriptions were detected with potential DDIs; the incidence was the highest; patients whose age over 65 y/o were the major population of potential DDIs. Of these 2,307 potential DDIs prescriptions, the severity of the potential adverse effect was rated as 1st grade in 12.9%, 2nd grade in 51.4%, 3rd grade in 4.7%, 4th grade in 14.9% and 5th grade in 16.1% respectively. A “serious” interaction which was defined as potentially life or organ threatening (1st grade) or disease aggravation (2nd grade) were estimated in a total of 64.3% potential DDIs. We found that digitalis Glycosides and loop Diuretics were the most common (28.16%) combination involving potential DDIs. From the results of decision tree, C5.0 algorithm was better than CART algorithm in terms of predictive accuracy; the first judgment of C5.0 model was principal diagnosis indicated its importance to which levels of DDIs was influenced. This research suggests that hospitals should set up clinical expert group to build a customized DDIs database for improving DDIs screening software systems.
