指導教授：魏志平臺灣大學：資訊管理學研究所謝采璇Hsieh, Tsai-HsuanTsai-HsuanHsieh2014-11-292018-06-292014-11-292018-06-292014http://ntur.lib.ntu.edu.tw//handle/246246/263503藥物不良反應是一個相當嚴重的全球性醫療問題，造成許多病人必須再度就醫，更嚴重的情況甚至導致死亡。因此，近年來受到許多國家的重視，進而推動藥物上市後的監控。起初，藥物上市後的監控是利用藥物不良反應通報系統來進行分析，但藥物不良反應通報系統是一個由醫護人員或是病人自行通報的系統，因此資料品質較難維護，並且可能有個人偏誤的報告。而後藥物上市後的監控轉向分析電子健康紀錄(Electronic Health Records)，因為電子健康紀錄包含病患的電子病歷或健康保險申報相關之就診資料，其為持續性的資料，涵蓋的人口範圍也較廣，資料內容亦較準確。 台灣的健保資料庫是一個健康保險申報相關之資料庫，本研究以台灣健保資料庫來進行藥物副作用的分析與探測，希望能利用學習排序法將藥物和可能導致的不良反應(疾病)的關聯進行排序，找出可能的藥物不良反應。我們建立四個實驗情境以評估本研究所提出之方法效能，其結果顯示，本研究提出之方法能夠有效的提升探測藥物不良反應的準確度，可以提供給專業的醫藥學專家進行進一步的驗證與分析。Pharmacovigilance (PhV) is a serious issue worldwide, because adverse drug effects are serious problems that cause harms to patients or even death. Traditionally, PhV research focuses on detecting adverse drug effects from spontaneous reports systems (SRS), which contains reports voluntarily reported by medical professionals, patients, and pharmaceutical companies. However, the volunteer nature of SRS databases causes some limitations (e.g., overreporting, data incompleteness). Thus, the PhV research starts to investigate the use of electronic health records (EHR) databases for drug safety signal detection in recent years. In this study, we propose a novel EHR-based drug safety signal detection method on the basis of the learning to rank approach. In addition to multiple disproportional analysis measures, our proposed method also incorporates as additional ranking variables that capture implicit relations between drugs and diseases for decreasing the importance of non-drug-outcome signals. We use Taiwan’s national health insurance research database for drug safety signal detection. Our evaluation results suggest that our proposed method significantly outperforms existing disproportional analysis methods (each of which uses a single disproportional analysis measures).誌謝 i 中文摘要 ii ABSTRACT iii List of Figures vii List of Tables x Chapter 1 Introduction 1 1.1 Background 1 1.2 Research Motivation and Objective 4 Chapter 2 Literature Review 7 2.1 Spontaneous Reports Systems (SRSs) 7 2.1.1 Definition and famous examples of SRSs 7 2.1.2 Methods used on SRSs 8 2.1.3 Traits of SRSs 10 2.2 Electronic Health Records (EHR) databases 11 2.2.1 Definition and famous examples of EHR databases 11 2.2.2 Methods used on EHR databases 12 2.2.3 Traits of EHR databases 13 2.3 Research Gap 14 Chapter 3 Design of Our Ranking Method 18 3.1 Data Collection 19 3.2 Data Preparation 20 3.2.1 Data Preprocessing 21 3.2.2 Drug-Appearing Diagnosis (DAD) Generation 24 3.3 Learning System 25 3.3.1 Drug or Disease Group Mapping 26 3.3.2 Labeling Signals 26 3.3.3 Measure Calculation for Training Data 27 3.3.4 Summery of All Measures 34 3.3.5 Ranking Model Building 35 3.4 Detection System 37 3.4.1 Drug or Disease Grouping and Signal Generation 38 3.4.2 Measure Calculation for Candidate Signals 39 3.4.3 Rank Prediction 39 Chapter 4 Evaluation and Results 40 4.1 Experimental Data 40 4.2 Evaluation Design 44 4.2.1 Evaluation Criteria 44 4.2.2 Evaluation Procedure 46 4.3 Comparative Evaluation 47 4.4 Additional Evaluation 48 4.4.1 Experiment 1: Effects of Variables Selection 48 4.4.2 Experiment 2: Effects of Training Sizes 52 4.4.3 Experiment 3: Effects of Surveillance and Control Window Sizes 57 4.4.4 Experiment 4: Appropriateness of Non-Mono-Domain Training 63 Chapter 5 Conclusion and Future Work 68 References 715342149 bytesapplication/pdf論文公開時間：2016/08/25論文使用權限：同意無償授權藥物不良反應資料探勘台灣健保資料庫thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/263503/1/ntu-103-R01725025-1.pdf