醫院急診部門壅塞現象將造成病患長時間的滯留與等待醫療資源，醫護人員若無法提供即時完善的醫療服務，將延誤治療時間並降低醫療品質，特別在急診室人口高峰時刻。由於急診室醫療資源的即時性十分重要，但因為過多的急診病患使得醫療資源管理變得相當複雜並提高其不確定性。有鑒於急診室壅塞已成為台灣地區十分嚴重且待解決的問題，研究取得合作醫院「台北馬偕教學醫院總院」之2010年急診室病歷，共計43,708筆資料以進行實證研究。本研究經與臨床醫師討論，首先欲瞭解造成急診室壅塞與長時間等待之病患診治流程與可能的行為模式。病患滯留時間(length of stay, LOS)在文獻與實務上都是急診室壅塞所關注的重要衡量指標，故本研究提出以病患滯留時間(LOS)為基礎，透過資料探勘相關技術，分析不同病患行為與LOS之關聯並進而建立急診室壅塞預測與分析模型。本研究建構以LOS為基礎的三階段急診室壅塞探勘與分析架構，嘗試探勘與分析不同LOS病患的顯著醫療行為，即顯著(regular)與例外(exception)病患的醫療行為。研究主要目的為(1)透過Apriori演算法切割不平衡資料集進而定義不同醫療行為，即顯著與例外病患；(2)使用K-Means分群法切割不同LOS的急診病患並分析不同LOS與急診室室壅塞趨勢之關聯；(3)基於分群結果，進一步使用分類法，如：J48、CART及JRip，建立LOS預測模型與萃取規則進行後續研究。本文所提出以Apriori演算法與分群演算法切割不平衡資料集進而找出顯著行為之輕症病患為急診室壅塞主要因素為研究方法重要的貢獻之一。研究透過資料探勘方法所找出之顯著病患醫療行為規則，經諮詢臨床醫師表示J48方法之規則結果有助於改善急診壅塞問題，所萃取的規則未來將結合決策支援系統，進而提供實務上的貢獻。由於台北馬偕教學醫院之急診室的就診人數為台灣第二大醫院，其急診資料在台灣地區具有一定程度的代表性，研究結果可提供醫院急診部門進行即時與完善之決策支援服務。

The emergency departments (EDs) of hospitals have been faced with the problem of access block or overcrowding, i.e., a long wait for medical resources or a high number of patients at the peak of time-delayed patient treatment, which may result in lower-quality healthcare. Because EDs must always be available to provide emergency medical care for patients, ED resource management can be extremely complex and uncertain. Accordingly, ED overcrowding is a national problem that requires a promising approach. In this research, we cooperate with the ED of the Mackay Memorial Hospital in Taipei to conduct an empirical study of a total of 43,708 records. Hospital ED crowding has led to increased patient wait times; thus, solving this problem requires a better understanding of patient flow and the behaviors of patients, according to the observations of clinical doctors in the ED. Furthermore, recent researches have highlighted the importance of analyzing the length of stay (LOS) in order to understand the behaviors of patients in the ED. They provides a good departure point for understanding patient behaviors based on the LOS factor. The aim of this study was to identify possible solutions for ED overcrowding in the Mackay Memorial Hospital in Taipei. Accordingly, we analyze the behaviors of patients with various LOS, and then predict the crowding state based on a data-mining technique in this research. We aim to propose a three-tiered, hybrid data-mining approach for mining ED patient behaviors from the set of patients' attributes and treatment processes (e.g., arrival to admission, examination pattern, transfer or discharge, etc.). In summary, the objectives of this research are briefly addressed here. (1) We applied association rule mining to identify frequent ED behaviors of patients (PBs) and infrequent ED behaviors of patients, i.e., the common and the exceptional behaviors of patients. (2) We adopted a K-means clustering approach in order to classify two types of PBs based on different LOS, and then labeled the cluster results using linguistic terms, i.e., long, medium, or short LOS. Furthermore, we conducted correlation analysis between various groups of LOS and the ED crowding condition. (3) We adopted classification techniques, i.e., J48, CART, and JRip, to build the LOS prediction model and extract rules for further investigation. To the best of our knowledge, the idea of partitioning the unbalanced data set by the Apriori algorithm and clustering techniques for investigating different users' behaviors in medical domains have not been addressed. Finally, we consulted with clinical doctors in order to confirm the results and rules extracted by J48 model that will be helpful for investigating the phenomenon of ED crowding in Taiwan, and we sought to build a medical-decision support system in the future. The Mackay Memorial Hospital in Taipei is the second-largest hospital in Taiwan; thus, it is a representative facility in Taiwan. Accordingly, the research results can serve as a reference for EDs for investigating ED crowding problems and making just-in-time decisions.