巨量醫療資料之時序事件追蹤與分析

Abstract

研究背景–全球高齡化趨勢以及社會型態的改變，使得人口健康問題與健康照護的支出日漸沉重。為防患於未然，各地的政策制定者紛紛推動醫療資料的電子化，用以幫助實現醫療保健系統的五大目標：(1)提高醫療的質量、安全及效率、(2)致力於病人所需要的健康照護、(3)增進健康照護的協調、(4)提高人口的健康、以及(5)確保私密性和安全性。然而，這些電子醫療資料要能“有意義的使用”、擴大用途、以及創造出更多效益，並不容易，且存在很多困難的研究議題待克服。目前醫療資料分散在不同業界，資料彙集困難，也鮮少相互連結分析。而且，醫療紀錄經長年累月已累積形成巨量資料(Big Data)，除對原有的計畫與研究造成重大衝擊，也帶來巨量醫療資料的整合、處理、分析、以及健康照護促進等新的研究議題，且衍生加值創新應用與商機。 研究目標–有鑑於目前生醫領域在巨量資料分析的基礎建設仍嚴重落後於趨勢、研究人員依然花費大量時間在建構與組織他們的資料，以及在這些資料上詮釋意義與發掘問題。為推動生醫巨量資料分析的變革，本研究的目標–提出一套從資料儲存到分析處理的完整方法，並基於此方法，驗證二個本研究所提出的巨量資料創新應用：(1)快速檢驗醫療通報事件，如藥物不良反應通報事件、(2)時序性醫療事件的及時監視追蹤，如新上市藥品的監視。為達成此目標，本研究提出的方法須具備：(1)時效性，能迅速回應處理結果、(2)效能，須以低成本達成、(3)擴充性，須能水平擴充運算能力及儲存容量、(4)計算容易，方便檢驗及追蹤指標的計算、以及(5)應用性。 研究方法–有別於流行病學研究方法，時序性醫療事件的追蹤與分析通常無法事先設定所要研究的問題。本研究提出一個新的模式，提供一個可及時追蹤監視醫療事件與揭露相關資訊的運作機制。此模式包含四個部分，分別為：(1)來源資料，即現有的電子醫療資料、(2)資料管理，包含巨量醫療資料儲存模型(PDMdoc)、時序性醫療事件模型(TMEdoc)、以及叢集的分片策略與管理等、(3)處理與運算，包含分片叢集的運作程序、雲端運算MapReduce巨量資料處理方法、以及一個整合的時序事件追蹤分析方法、(4)追蹤指標，內容包含指標項目，以及記載每一個發生此事件的病患的指標數值。其中，影響此模式能否發揮及時監視追蹤的功能，關鍵在於資料管理以及處理與運算的效能。 結果–本研究方法的複雜度：(1)叢集的水平擴充性與平行度為1，亦即每加入一個分片節點至叢集系統，運算能力及儲存容量均會增加一個分片單位，且不受叢集的節點數影響、(2)網路I/O，僅與查詢結果的資料量有關，也不受節點數影響、(3)搜尋與磁碟I/O，PDMdoc與TMEdoc平均搜尋時間分別為O(1)與O(logd(STMEdoc/B))，平均磁碟I/O (尋軌時間、旋轉延遲、傳遞時間)分別為( O(1), O(1), O(EPDMdoc) ) 與 ( O(logd(STMEdoc/B)), O(1), O(ETMEdoc × LTMEdoc) )。在實驗方面：(1)資料，取自於全民健保資料庫承保抽樣歸人檔(LHID2010)，計有100萬人從1996年至2010年期間的健保資料、(2)測試系統，採用MongoDB以及5部PC建置成分片叢集系統(3個分片節點)、(3)實驗結果：(a)效能測試，搜尋8個疾病族群的罹病患者，單伺服機系統與分片叢集的花費時間分別介於0.607∼63.248與0.336∼29.484秒，二者平均效能比為1 : 2.024、(b)藥物不良反應通報事件的檢驗，以美國FDA於2009年9月發布的Januvia藥品安全資訊為通報案例，檢驗結果(odds ratio = 1.626)顯示此事件在台灣也有顯著的情形、(c)新上市藥品的監視，系統處理TME的數量可達140,000/秒以上，估計每日可監視數千至數萬種藥品。

Backgroud – Global aging trend combined with societal changes are creating population health problems and increasing health care spending. As a precaution, local policy makers have been promoting electronic medical data to help achieve five major goals of health care system: 1) improving health care quality, safety, and performance, 2) committing to patient health needs, 3) improving health care coordination, 4) improving the health of the population, and 5) ensuring privacy and security. However, in order to make these medical data to be 'Meaningful Use', to expand data usage, and to create more profits, many research difficulties have to be overcome and it will not an easy task. Currently medical data is scattered in different industries, data collection is difficult, and mutual analysis is rare. Furthermore, medical records have been accumulating to big data after many years. This not only significantly impacts original plan and research, but also creates bonus innovative applications and opportunities. Objectives – Given that the current biomedical field in big data analysis infrastructure is still seriously lagging behind current trend, researchers have to spend considerable time on constructing and organizing their data and on interpreting meaning and identifying issues with these data. To revolutionize biomedical big data analysis, this study proposes a set of methods ranging from data storage to data analysis. Based on this set of methods, two novel applications for big data were verified, 1) prompt testing of medical reported incidents, such as adverse drug reactions reported incidents, 2) timely monitoring and tracking of temporal medical events, such as monitoring of newly marketed drugs. To achieve the objectives, this set of methods must have: 1) timeliness, to quickly respond process results, 2) effectiveness, shall reach low cost reach, 3) scalability, shall allow horizontal expansion of computing power and storage capacity, 4) easy calculation, convenient for testing and calculating tracking indicators, and 5) applicability. Methods – Unlike epidemiological research methods, problems to be studied for tracking and analysis of temporal medical events cannot be delivered in advance. This study proposes a new model, providing an operation mechanism which allows for timely tracking and monitoring of medical events and uncovering relevant information. This model contains four parts, which are: 1) source of data, namely current electronic medical data, 2) data management, including big data storage model PDMdoc, temporal medical events model TMEdoc, and tactics and management of sharded cluster, 3) processing and computing, including sharded cluster operating procedures, cloud computing MapReduce big data processing methods, and an integrated temporal event tracking analysis, 4) tracking indicator, content mainly comprising of a number of indicators, and recording patient index value for every occurrence. Among them, indicators belong to practical application level; therefore impacting whether this model can achieve timely monitoring and tracking function, the essential part lies in data management and efficiency of processing and calculation method. Results – Complexity of the research methods in this study: 1) sharded cluster horizontal scaling and degree of parallelism is 1 unit, specifically, every time a shard is added to the cluster system, the computing power and storage capacity will both be increased by 1 unit, not affected by the number of cluster nodes, 2) network I/O, only relevant to the amount of data for search results, irrelevant to the number of cluster nodes, 3) search and disk I/O, average seek time for PDMdoc and TMEdoc are O(1) and O(logd(STMEdoc/B)), respectively, average disk I/O for seek time, rotational delay, transmission time are 'O(1), O(1), O(EPDMdoc)' and 'O(logd(STMEdoc/B)), O(1), O(ETMEdoc × LTMEdoc)', respectively. Statistics in experiments performed, 1) data, gathered from Taiwan NHIRD LHID2010 Dataset, containing health care data of a total of one million people for the period 1996 to 2010, 2) test system, sharded cluster containing 3 shard nodes built on MongoDB and five PCs, 3) experiments results: a) benchmarks, the times required to search diseased patients from 8 disease groups for single server system and sharded cluster range from 0.607 to 63.248 seconds and from 0.336 to 29.484 seconds, respectively, the two systems have performance ratio of 1:2.024, b) adverse drug reactions reported incidents, take Januvia drug safety information published by FDA in September, 2009 for example, the test result for odds ratio is 1.626, showing that this type of incidents had significant occurrences in Taiwan as well, c) monitoring for newly marketed drugs, system processing capacity for number of TME can exceed 140,000 per second, the daily number of drugs that can be monitored is estimated to be above tens of thousands.