急診處抗藥性格蘭氏陰性菌血症預測模型之建立

Abstract

前言: 格蘭氏陰性菌血症是臨床上常見且預後不佳的感染急症。抗生素抗藥性的出現使得這樣的情形更加惡化。針對高抗藥性的格蘭氏陰性菌，雖然已經有許多國外文獻試圖找出流行病學上各種的危險因子，但是對於將危險因子量化以直接協助臨床應用的研究仍付之闕如，而有關台灣格蘭性陰性菌抗藥性危險因子的本土性資料亦少有以英文發表之研究報告。

目的： 找出具有抗藥性格蘭氏陰性菌血症病患之相關危險因子及各種臨床特徵，並嘗試以此流行病學之量性結果建立預測模式，以提供臨床醫師在細菌鑑定結果及抗生素感受性確認前，治療疑似格蘭氏陰性菌感染症病人時的藥物選擇參考。

材料與方法: 自民國 90 年 6月1日起至民國91年5月31日止，以在台大醫院急診醫學部檢驗出的格蘭氏陰性菌血症病患之臨床資料進行之前瞻性研究。排除15歲以下的兒童及創傷病患。收集的資料包含病患背景、先前疾病、醫院接觸史(包括詳細的住院史及健康照護相關史如洗腎或門診化療等) 及到院之臨床表現等為變項 (exposures)。定義以針對第一代環孢靈素 (以cefazolin 為代表; 簡寫為 CZ-RES) 出現抗藥性及針對第三代環孢靈素 (以ceftriaxone 為代表；簡寫為 CTX-RES) 出現抗藥性之格蘭氏陰性菌菌血症為事件(outcomes)。並以隨機分派 (random allocation) 的方式將所有資料的三分之二列入模式導出組 (derivation)，三分之一列入模式驗證組 (validation)。 將導出組中單變項分析發現有意義的結果導入羅吉斯氏迴歸 (logistic regression) 之多變項分析以建立統計預測模式。除了在驗證組測試預測模式之效度外，亦使用參考變項係數之整數化給分法(coefficient-based scoring method) 簡化預測模式以方便臨床應用。

結果: 研究期間共收集 695筆格蘭氏陰性菌血症之病人資料。針對 CZ-RES 預測模式，經羅吉斯氏迴歸求得之危險因子為「本次菌血症距離前次出院時間」、 「前次住院曾有感染對ceftriaxone 有抗藥性之細菌」、 「移植後正在服用免疫抗制劑的病患」、 「病患來源是否為安養院或病患本身是否為中風併有反覆嗆入史」 及「病患到達急診處時的血氧濃度小於 95%」。在本研究中，「肝硬化」 與感染抗藥性格蘭氏陰性菌血症之機會呈現逆相關。以上述因子建立對「CZ-RES」 之預測模式，其使用者操作特徵曲線 (Receiver Operating Characteristic curve; ROC curve) 下面積為 0.76，其 95% 信賴區間為0.71 ~ 0.81。 針對 CTX-RES 預測模式，選入的變項除包含了所有 「CZ-RES」的預測因子外，尚包括的危險因子為「病患到達急診處第一次驗血的白血球數目不正常( <1000/mm3 或 > 15,000 /mm3)。」以上述因子建立對「CTX-RES」 之預測模式，其使用者操作特徵曲線 (ROC curve) 下面積為 0.82，其 95% 信賴區間為0.76 ~ 0.88。 簡化後的參考係數之整數化給分預測模式，其 ROC curve 下面積與原導出模式非常接近。

結論: 本研究以台灣本土急診病患為對象，找出在可能產生抗藥性格蘭氏陰性菌血症之危險因子並具以建立有效度驗證之預測模式。此預測模式更被簡化成危險因子整數分數計分以方便臨床使用。在第一線醫師面對可能患有格蘭氏陰性菌血症感染的病人、且細菌鑑定及抗生素感受性報告尚未得知時，此預測模式將有助於正確地選擇經驗性抗生素。

Background The increasing prevalence of antimicrobial resistance among gram-negative bacteria has increasingly gained attention. Despite numerous studies on risk factors related to gram-negative antimicrobial resistance, there was short of predictive model underpinning quantitative epidemiological findings, particularly in Taiwan, for the prediction of antimicrobial resistant gram-negative resistance before bacterial culture result is released.

Objectives To find out the risk factors for gram-negative resistant bacteremia in Taiwan and to develop a predictive model to assist physician in appropriate selection of the empirical antimicrobial agent before the microbiologic idenditification and drug susceptibility known.

Material and Methods A prospective study was conducted form June 1, 2001 to May 31, 2002 at emergency department (ED) in National Taiwan University Hospital. Enrollees were patients with gram-negative bacteremia sampled at ED. Collected exposures included demographic characteristics of patients, underlying comorbidities, hospital exposure and health-care associated factors, and initial presentation. Two primary outcomes were defined as cefazolin-resistant gram-negative bacteremia (CZ-RES) and ceftriaxone-resistant gram-negative bacteremia (CTX-RES). Two-third of data was randomly allocated to a derivation dataset for training parameters pertaining to predictive models and the others to a validation dataset for testing model validity. Simplified models by coefficient-based scoring method were also established for ease of clinical application.

Results There were total 695 episodes of gram-negative bacteremia in final analysis. Predictors identified for CZ-RES gram-negative bacteremia included length from prior hospitalization to existent bacteremia (increasing risk within one month), prior infection by ceftriaxone resistant strain, post-transplantation patients with immunosuppressant in use, nursing home residence or history of cerebral vascular accidents with repeated chocking events, and poor oxygen saturation (<95%) at arrival at ED. Cirrhosis showed its protective effect in reducing the odd of antimicrobial resistant gram-negative bacteremia. As to CZ-RES models, the area under receiver operating characteristic curve (ROC curve) was 0.76 (95% C.I.: 0.71 ~ 0.81)(C.I.: confidence interval). The CTX-RES model included all predictors in CZ-RES model together with abnormal leukocyte count (<1000 or > 15,000 /mm3) at first blood sampling at ED. Besides, the risk temporal length form prior hospitalization is shorter (increasing risk within two weeks). The area ROC curve was 0.82 (95% C.I.: 0.76 ~ 0.88). Area under ROC curve of two simplified integral scoring models was very close to the models by derivation sets.

Conclusion We developed two quantitative predictive models by the application of identification and quantification of risks factors associated with antimicrobial resistant gram-negative infection. Application of these predictive models provided in this study can help physician in choosing empirical antibiotic appropriately before the bacterial culture result available.