手術傷口感染判別模式之建立與應用：以醫療提供者冠狀動脈繞道術服務量與感染研究為例

Abstract

背景 院內感染是當今醫療照護的一大問題，學界在過去也進行過許多大型研究，來釐清影響院內感染發生的因子，服務量-感染研究即是其中一例。然而，在過去的研究中，服務量與感染之間的關係尚無法獲得釐清。可能的原因包含服務量的定義不一、分析方法的限制以及判定感染個案方法的正確性。因此，本研究之目的在於發展適合台灣健保資料庫使用之CABG手術傷口感染判別模式，並將所發展之判別模式，應用至全民健保資料庫，以進行後續服務量-感染研究，探討醫院、醫師服務量與手術傷口感染之關係。 研究設計 回溯性研究 材料與方法 藉由文獻回顧，本研究收集國際間近年來利用申報或行政資料所發展感染判別模式。並利用某醫學中心2005-2008冠狀動脈繞道術病人申報資料，以及院內感控監視系統，進行各種冠狀動脈繞道術手術傷口感染個案判別替代模式的建構（alternative models），並進行傳統模式（ICD-9 CM）與其他替代模式的比較。在將所得之最佳模式，應用至2005-2009年間全國冠狀動脈繞道術申報資料，進行感染個案判別，以及後續服務量-感染分析。在服務量感染分析部分，本研究將服務量分為醫院/醫師當月短期服務量以及手術前12月長期累積服務量，並採用廣義加法模式、k-means分群法以及四分位法進行服務量分群，以多層次分析方法，探討醫院與醫師服務量與手術傷口感染之關係。 結果 本研究發現，決策樹模式有最佳的陽性預測值（PPV=77.78%），而且在敏感性、特異性與陰性預測值的表現也十分突出。此外，本研究也發現，幾乎各種判別模式在陽性預測值的表現，都比使用ICD-9的傳統模式為佳。在考量資料同質性後，本研究僅就2006-2008年醫學中心健保申報資料，進行感染個案判別與分析。本研究發現，經判別模式判別後，2006至2008年間醫學中心預測感染率約1.5%，病患的抗生素與醫療利用有逐年減少的趨勢。本研究也發現，使用不同的服務量分類方式，會有不一樣的結果。而使用廣義加法模式之分類方式，可以取得最佳之模型適配度（AIC=1083.34），結果顯示，醫師的前12月累積服務量越多，病患感染風險越低（OR=0.9888）。但在醫院服務量的部份，則是呈現不論是當月或累積服務量，中服務量的醫院，感染風險比低服務量與高服務量醫院來得高，分別為1.9058與1.7510。 結論 相較於申報資料中的ICD-9 CM碼，透過決策樹判別模式，可以較精確的找出感染個案，幫助研究者減少研究上的偏誤。而服務量與感染的關係尚未明確，本研究以各種不同服務量分類方式進行分析與比較，可作為後續研究之參考

Background Hospital-acquired infections are an important issue in the current health care society. Several large-scale studies have been conducted to verify factors that could affect hospital-acquired infections, including those where the relationship between service volume and infection was examined, although it remained controversial. Possible reasons for this equivocal relationship include inconsistent definition of service volume, limitations in data analysis methods, and inappropriate approaches adopted for identifying infection cases. Therefore, the purpose of this study is to develop a model which can be applied in the National Health Insurance (NHI) Research Database of Taiwan for identifying surgical site infection following CABG surgery. The model is further applied in later studies to examine the relationship between hospital and surgeon services volume and surgical site infection. Study design A restrospective study design. Materials and method Through literature review of domestic studies and those from abroad, the study collected models for infection identification which were developed upon registry or administrative data. Using the registry data of CABG surgeries between 2005 and 2008 from a medical center in Taiwan and information collected by the hospital surveillance system, the study established an alternative model for identifying surgical site infection following CABG surgery. This model was compared with the traditional model based on ICD-9 CM codes and other alternatives. Then, the derived optimal model was applied in the registry data of CABG surgery between 2005 and 2009 in Taiwan for infection case identification. The model was also used in later studies for exploring the service volume-infection relationship. For statistical analyses, service volume was divided into shorten monthly hospital/surgeon services volume and long-term cumulative services volume during the past 12 months of a surgery. The generalized additive model, k-means clustering, and quartiles were used for stratifying volume groups, and multi-level analysis was performed to examine the relationship between hospital and surgeon services volume and surgical site infection. Results The study results revealed that the decision tree model owned the highest positive predictive value (PPV) of 77.78 % and was qualified with good sensitivity, specificity, and negative predictive value. Furthermore, the study found that almost every alternative model had a PPV better than that of the traditional ICD-9-based model. Considering the homogeneity of data, the study only conducted its identification and analysis of infection cases in the NHI registry data of medical centers from 2006 to 2008. The study results showed an infection rate of 1.5 % in medical centers between 2006 and 2008, accompanied with a decreasing trend of antibiotics and healthcare services utilization. The study also found that different categorizations of service volume yielded different results, where a categorization using the generalized additive model brought the best goodness of fit of the model (AIC=1083.34). The study results suggested that greater cumulative physician services volume during the past 12 months of a surgery led to reduced risk of infection in patients (OR=0.9888). Regarding hospital services volume, no matter monthly or cumulative volume was adopted, middle-volume hospitals showed higher risk of infection in patients than low-volume and high-volume hospitals, where the OR were 1.9058 and 1.7510, respectively. Conclusion Compared to ICD-9 CM codes in claim data, a decision tree-based model can identify infection cases more precisely, which reduces bias in studies. When conducting studies regarding service volume and infection, the definition and categorization of service volume merit concern as results could vary accordingly. The study analyzes and compares among different categorizations of service volume, and the findings could serve as references for future studies.