應用以網格運算為基礎之免疫演算法改善高維度分數法於干擾控制之研究

Use of High-Dimensional Propensity Score with Grid Computing based Immune Algorithm to Improve Confounding Control

李宜澤(Yi-Che Lee)

指導教授 ： 陳大正

國立虎尾科技大學/管理學院/資訊管理研究所/碩士(2014年)

摘要

本研究利用台灣健保資料庫下，欲探討比較性的治療與效用(effectiveness)之研究時，如何於巨量資料(big data)中探勘出最具潛力的干擾共變數組合，藉此來修正干擾共變數所產生的偏差(Bias)而達到最佳的干擾調整。
近年來許多學者在藥物流行病學的研究中指出，應用以高維度傾向分數(hd-PS)為基礎的方法時，若能適當運用豐富資料量的巨量資料庫進行研究，所獲得的調整偏差來估算，則其結果與隨機試驗及觀察性研究會非常接近，能比其他方法獲得更佳的干擾調整。然而藉由該法如何由巨量資料中探勘出最具潛力的干擾共變數仍是過去文獻尚未能突破之處。因過去應用hd-PS於選擇潛在干擾共變數的過程中需高度仰賴由研究者或文獻所提供的專業知識及經驗。
透過本研究，提出一進化演算法，將hd-PS方法延伸轉換為一個於巨量健保資料庫中，進行重要干擾共變數的探勘，達成干擾調整的最佳化問題。因此本研究將提出一個以網格運算為基礎之進化演算法改善hd-PS方法來求取最佳干擾共變數之組合。研究結果顯示，本研究所提方法，所探勘出的潛在干擾共變數之組合，比過去hd-PS方法較易探勘出潛在的干擾共變數組合，治療與效用之研究達到勝算比最小化。

關鍵字

巨量健保資料庫 ； 比較效果研究 ； 高維度傾向分數 ； 傾向分數 ； 網格運算

並列摘要

Through Taiwan’s Health Claim Database, this study investigated the confounding variable combination with the best potential in the big data based on comparative studies on therapy and effectiveness, thereby modifying bias arising form the confounding covariates and achieving the best confounding factor adjustment. In recent years, a number of scholars have pointed out in their research on drug epidemiology that when high-dimensional propensity score-based (hd-PS) approach is applied, through the proper use of rich data in the big database for research and the acquisition of adjusted bias for estimations, the results will be very close to stochastic testing and observational studies, thus deriving at better confounding factor adjustment compared to other methods.
However, how to explore confounding covariates with the highest potential from the big data remains to be a bottleneck faced in past literatures. In the past, the process of selecting confounding covariates using hd-PS involved high reliability on professional knowledge and experiences provided by researchers or literatures.
Through this study, an evolutionary computation approach was proposed to covert the hd-PS method in the Health Claim database with a massive amount of data for exploring important confounding covariates and achieving the optimization of confounding factor adjustment. Hence, a grid computing based evolutionary computation approach was proposed to improve the hd-PS method and obtain the best confounding covariate combination. Findings show that the method put forth in this study was better able to explore the potential confounding variable combination compared to past hd-PS methods, thus achieving the minimization of odds ratio in therapy and efficacy related studies.

並列關鍵字

Big Data ； Comparative Effectiveness Research ； High-Dimensional propensity score ； propensity score ； Grid computing

參考文獻

[48]曹惠鈴，2007，“運用混合式進化演算法於法則探勘之回應模型”，國立虎尾科大資料管理系，碩士論文。

[50]尤婷藝，2010，“混合式進化演算法於微陣列資料分類法則探勘之研究”，國立虎尾科技大學資訊管理系碩士論文。

[1]Hunt, J. E. and D. E. Cooke, (1996). , "Learning using an artificial immune system, " Journal of Network and Computer Applications, Vol. 19, pp. 189-212.

[2]Bross, I. D. (1966). Spurious effects from an extraneous variable. Journal of chronic diseases, 19(6), pp. 637-647.

[3]Bombardier, C., Laine, L., Reicin, A., Reicin A., Shapiro, D., Burgos-Vargas, R., Davis, B., Day, D.,Fe1Taz, M.B., Hawkey, C.J., Hochberg, M.C., Kvien, T.K., and Schnitzer, T.J. (2000), "Comparison of upper gastrointestinal toxicity of rofecoxib and naproxen in patients with rheumatoid artluitis," The New England Journal of Medicine, 343, pp. 1520-1528.