巨量資料之病例對照研究平台

Abstract

本研究的目的在於設計一套流程，用於產生觀察型研究的研究對象，並且實作系統。透過資料篩選、變數的建立與處理、對照組的匹配以及初步的統計檢定四個步驟，讓使用者對研究對象有一個概觀，並得到資料的雛型，以做進一步的研究。 我們以國家衛生研究院發行的百萬人承保抽樣歸人檔(LHID2010)作為研究資料。這份資料包含了台灣2010年全民健保在保者的100萬人抽樣檔，其在1996年至2010年間的所有就醫資料，以歸人的形式建立。這樣的資料結構與龐大的資料量正適合利用NoSQL資料庫schema free與水平擴充的特性來處理。 因此我們建立了MongoDB replica sharded cluster，利用分片(sharding)的功能，可以提升查詢效率，再配合Map-Reduce方法，可以對資料進行較複雜的運算，產生修整好的資料，提供後續的統計分析。

The objective of this research is to make a platform, helping users to obtain the study population of observational studies by a four stages procedure including data querying, variables creating, control group matching and significance testing. Data querying platform was used to find out the study population. It provides a simple interface for NoSQL database query. It also automatically makes a flow chart of query results, helping users manage the process of query. Variables creating platform let users extract detail attribute information of selected patients from previous stage such as disease diagnoses or drugs taken records. It was done by a Mongo Map-Reduce process and export to a csv file for next stage. Control group matching platform read the patients and variables from the previous stage and do propensity score matching. Users choose treatment (or exposure), outcome or other covariates from the variables, then fit the generalized linear model and match the control group by fitted value. Significant testing platform did t-test on each variable and chi-square goodness of fit test on each age group between case group and control group to see if there is any significant difference. The first two stages of working process can be separated from others into two independent parts. The first part prepares data for observational study, the second part implements statistic analyzing. Users may have their own analyses on the data we prepared or loading their own data into our matching platform.