Original Contribution

predictive model of antimicrobial-resistant Gram-negative bacteremia at the ED

Wen-Chu Chiang MD, MPH^a, Shey-Ying Chen MD^a, Kuo-Liong Chien MD, PhD^b, Grace Hui-Min Wu, PhD^b, Amy Ming-Fang Yen PhD^b, Chan-Ping Su MD^a,

Chien-Chang Lee MD^a, Yee-Chun Chen MD, PhD^c, Shan-Chwen Chang MD, PhD^c, Shyr-Chyr Chen MD^a, Wen-Jone Chen MD, PhD^a, Tony Hsiu-Hsi Chen DDS, PhD^b,*

^aDepartment of Emergency Medicine, National Taiwan University Hospital, Taipei 100, Taiwan

^bInstitute of Preventive Medicine, College of Public Health, National Taiwan University, Taipei 100, Taiwan

^cDivision of Infectious Disease, Department of Internal Medicine, National Taiwan University Hospital, Taipei 100, Taiwan

Received 17 August 2006; revised 30 November 2006; accepted 30 November 2006

Abstract

Background: Despite numerous studies identifying the risk factors related to Gram-negative Antimicrobial resistance, an epidemiological model to reliably predict antimicrobial Gram-negative resistance in clinics, before the Bacterial culture result is available, has not yet been developed.

Objectives: The aim of this study was to develop a predictive model to assist physicians in selecting appropriate antimicrobial agents before the details of the microbiology and drug susceptibility are known. Materials and Methods: A prospective study was conducted between June 1, 2001, and May 31, 2002, at the emergency department (ED) of National Taiwan University Hospital. Enrollees were patients with Gram-negative bacteremia (GNB) at ED. Other information collected included demographic characteristics, underlying comorbidities, hospital exposure and health care-associated factors, and details of initial presentation. Two primary outcomes were defined, including cefazolin-resistant (CZ- RES) GNB and ceftriaxone-resistant (CTX-RES) GNB. Two thirds of the data was randomly allocated to a derivation data set (for developing predictive models), and the rest, to a validation data set (for testing model validity). Simplified models, using a coefficient-based scoring method, were also developed for clinical applications.

Results: Based on 695 episodes of GNB, predictors of CZ-RES GNB were time since last hospitalization (increased risk for durations b1 month), prior infection with a CTX-RES strain, post- transplantation immunosuppressant use, residence in a nursing home or history of stroke with repeated choking, and poor oxygen saturation (b95%) at admission to ED. Cirrhosis showed a protective effect by reducing the odds of antimicrobial-resistant GNB. The area under receiver operating characteristic (ROC) curve for the CZ-RES model was 0.76 (95% confidence interval, 0.71-0.81).

The CTX-RES model included all the variables that were in the CZ-RES model plus abnormal

* Corresponding author. Tony Hsiu-Hsi Chen, Institute of Preventive Medicine, College of Public Health, National Taiwan University, Taipei 100, Taiwan. Tel.: +886 2 23516478 (25); fax: +886 2 23920456. Address for reprints: Wen-Chu Chiang, Department of Emergency Medicine, National Taiwan University

Hospital, Taipei 100, Taiwan. Tel.: +886 2 23562168; fax: +886 2 23223150.

E-mail addresses: [email protected] (W.-C. Chiang)8 [email protected] (T. Hsiu-Hsi Chen).

0735-6757/$ – see front matter D 2007 doi:10.1016/j.ajem.2006.11.024

leukocyte count (b1000 or N15000 /mm³) at entry to ED. In this case, however, previous hospitalization within the last 2 weeks was a key factor. The area under this ROC curve was 0.82 (95% confidence interval, 0.76-0.88). There was lacking of difference in the area under the ROC curve between the 2 final (simplified) models either based on the derivation or validation data sets.

Conclusion: We have developed 2 models for predicting risk of antimicrobial Gram-negative infection by identifying and quantifying associated risk factors. These models could be used by physicians to determine the most appropriate choice of antibiotic for first-line therapy, particularly in situations where the culture result is not yet known.

D 2007

Introduction

Gram-negative bacterial infections are becoming increas- ingly prevalent in many populations and have, because of their association with unfavorable clinical outcomes, attracted much attention over the past decade. In particular, such infections often lead to longer hospital stays, increased attributable mortality, and greater Hospitalization costs [1-5]. The identification of risk factors affecting susceptibility to antibiotic resistant Gram-negative pathogens is therefore of paramount importance.

Risk factors identified in previous studies included demographic characteristics, underlying comorbidities, and the type or pattern of medical care. Subjects who are older than 65 years have been reported as being a high-risk group [6], although few studies have corroborated this. Having hepatic failure or end-stage renal disease [7], or having had gastrointestinal (GI) surgery or an organ transplant [8-10] makes infection with an antimicrobial- resistant Gram-negative pathogen more likely. Although diabetes mellitus is more susceptible to infection, it was not associated with antimicrobial resistance in many reviews [6-16]. Hospital exposure and health care-associ- ated factors were investigated in many studies, more specifically, long-term hemodialysis at clinics or hospital [11,12], residence in a nursing-home [12,13], history of prolonged hospital stays [8,9,14] having an indwelling catheter (such as central venous, arterial [7,13], or urinary) [7,15,16], requiring mechanical ventilation [7,15,16], and being tube-fed [10]. Previous anti-infective treatment with fluroquinolones [8,9] and multiple antibiotic exposure lasting longer than 14 days [6,10] were also considered as risk factors.

There are 3 deficiencies in these previous findings. Firstly, most of the research was based on retrospective studies that used a slightly different definition of antibiotic resistance, making the results difficult to interpret. Secondly, with so many risk factors identified, a physician may be faced with an impossible task when trying to choose an appropriate antibiotic. Finally, as antimicro- bial resistance develops through the use of antibiotics, which varies from country to country, the findings from studies carried out in the west may not be generalized to other populations.

As cases of antimicrobial-resistant Gram-negative bac- terial infection differ considerably as regards host factors, the objective of this study was to use background patient information to develop a series of predictive models that could help physicians identify high-risk patients and decide which antimicrobial agent to prescribe.

Materials and methods

Design, setting, and enrollees

Patients with bacteremia were identified prospectively during the period June 1, 2001, through May 31, 2002, from attendees at the emergency department (ED) of National Taiwan University Hospital (NTUH) in Taipei, which has a capacity of 2400 beds and more than 100000 ED visits annually. Note that only nontrauma adult patients (ie, N15 years old) with proven Gram-negative bacteremia (GNB) were enrolled in this study. The details of study process from initial report of bacterial growth to the identification of exposure and outcome assessment are summarized in Fig. 1. This study was approved by the institutional review board of NTUH.

Cephalosporins were extensively used to treat Gram- negative bacterial infections in Taiwan [17]. First-generation cephalosporin remains sensitive to most Gram-positive and Gram-negative pathogens. When the in vitro test shows susceptibility, they are a better choice because they are inexpensive and induce resistance less than the second- and third-generation cephalosporins [18,19]. In contrast, third- generation cephalosporin has the widest spectrum of Gram- negative activity and is preferred in most serious Gram- negative infection. Furthermore, if the pathogen is suspected of being nonfermenting, Gram-negative bacilli or extended- spectrum b-lactamase Enterobacter spp, a stronger antimi- crobial agent, such as imipenem, will be prescribed.

Exposures assessment

To develop a predictive model for antimicrobial-resistant bacteremia, we collected information on as many of the previously identified predictors of antimicrobial-resistant Gram-negative bacterial infection as possible. Table 1 specifies 60 items of exposures, including 4 blanket terms:

Fig. 1 Flowchart of study design.

demographic characteristics, underlying comorbidities, hospital exposure and health care-associated factors, and initial presentation. Medical history and clinical presenta- tions of all eligible patients were obtained by person-to- person interview and detail medical chart review by the stationed physician. During this evaluation stage, the physician who carried out all the questionnaire interviews and chart reviews was unaware of the details of the infection and blind to the outcome of antibiotic resistance because the final microbiology report would not have been available until 3 to 5 days after the first laboratory telephone report.

Outcome management

We set out 2 prediction models to assist physicians’ in choosing antibiotics. The first model is related to the resistance to first-generation cephalosporin, defined as cefazolin-resistant (CZ-RES) GNB, and the second is pertinent to the resistance to third-generation cephalosporin

(without anti-Pseudomonas activity), defined as ceftriax- one-resistant (CTX-RES) GNB. Data relating to bacteremic episodes caused by Salmonella spp, Aeromonas spp, Vibrio spp, or Campylobacter spp were excluded from analysis because the clinical features, predominant risk factors, and prescribing practice with such infections were largely different from those of the most common Gram-negative infections [20-25].

Microbiological assessments and antimicrobial susceptibility

Blood from patients was stored in BACTEC standard or BACTEC PLUS culture bottles in a BACTEC 9000 system (Becton Dikinson, Cockeysville, Md) for blood culture. Identification of the isolates and susceptibility testing was performed at the central laboratory in NTUH using standard bacteriologic techniques and an automated system (VITEK, bioMerieux Inc, USA).

Statistical method for predictive models

We adapted derivation-validation method to develop predictive models. Detailed processes of the enrollment and randomization are given in Fig. 2. Continuous data were categorized by splitting into groups with appropriately chosen cut points. Univariate associations between categor-

Fig. 2 Enrollment, data collection, and assignment.

Table 1 Univariate analyses of derivation data set (n = 460)

Exposure: E (+)	Total E (+)	CZ-RES				CTX-RES
		Cases	OR	95% CI		Cases OR	95% CI
demographic factors
Sex
Male	230	74	1.5*	1.0-2.3		35 1.3	0.8-2.2
female	230	55	1.0	–		28 1.0	–
Age (y)
V40	53	15	1.0	–		7 1.0	–
41-64	145	35	0.8	0.4-1.6		22 1.2 0.5-2.9
65-75	128	37	1.0	0.5-2.1		14 0.8	0.3-2.1
z75 Comorbidity factors	134	42	1.2	0.6-2.3		20 1.2	0.5-2.9
Diabetes mellitus	139	38	1.0	0.6-1.5		18 0.9	0.5-1.6
Liver cirrhosis	48	6	0.3*	0.1-0.8		2 0.3	0.1-1.1
End-stage renal disease	11	4	1.5	0.4-5.2		2 1.4	0.3-6.7
Stroke with repeat choking history	52	27	3.2.**	1.8-5.8		16 3.4**	1.8-6.6
Bedridden
Bedridden, no bedsore	46	22	2.7**	1.5-5.1		13 3.0**	1.5-6.2
Bedridden, with bedsore	13	6	2.6	0.8-7.8		4 3.4*	1.0-11.6
alcohol consumption	7	1	0.4	0.1-3.6		0 –	–
Long-term steroid needed	14	4	1.0	0.3-3.3		2 1.1	0.2-4.8
Hematological malignancy	25	15	4.2**	1.9-9.7		8 3.3**	1.3-7.8
Solid organ malignancy
GI tract malignancy	62	19	1.7	0.8-3.9		11 4.1**	1.7-9.7
Non-GI tract malignancy	26	10	1.2	0.7-2.2		9 1.7	0.8-3.4
Chemotherapy	36	18	2.8**	1.4-5.6		13 2.8**	1.4-5.6
Immunosuppressant	10	4	1.7	0.5-6.3		2 1.6	0.3-7.9
autoimmune disease	9	3	1.3	0.3-5.2		2 1.8	0.4-9.0
Health care-associated factors
Patient source
Community	390	91	1.0	–		42 1.0	–
Nursing home	16	10	5.5**	1.9-15.5		5 3.8*	1.3-11.4
Long-term health care needed	30	11	1.9	0.9-4.2		4 1.3	0.4-3.8
Hospitalization in last month	24	17	8.0**	3.2-19.8		12 8.3**	3.5-19.6
Time since the last discharge (d)
b7	46	30	8.3**	4.2-16.5		21 10.9**	5.1-23.4
8-14	19	9	4.0**	1.5-10.4		6 6.0**	2.0-17.8
15-21	20	11	5.4**	2.1-13.8		1 0.7	0.1-5.4
22-30	19	8	3.2*	1.2-8.4		4 3.5*	1.1-11.6
31-45	18	4	1.3	0.4-4.0		4 3.7*	1.1-12.5
46-365	100	23	1.3	0.8-2.3		10 1.4	0.6-3.3
No hospitalization with the last year	238	44	1.0	–		17 1.0	–
Last admitted hospital
No hospitalization with the last year	238	44	1.0	–		17 1.0	–
Admitted, not at NTUH	37	21	5.8**	2.8-12.0		12 6.2**	2.7-14.6
Admitted at NTUH	165	64	2.3**	1.5-3.6		34 2.9**	1.6-5.4
Length of last hospital stay (d)
No hospitalization within last year 238		46	1.0	–	17	1.0	–
b7 54		22	14.1**	4.9-40.6	12	14.3**	5.3-38.4
8-15	65	22	3.4**	2.1-5.5	11	3.8**	2.0-7.3
16-30	58	23	1.1	0.5-2.3	13	1.1	0.4-3.4
N30	43	16	1.3	0.5-3.3	10	2.0	0.6-6.4
Cause of prior hospitalization
No hospitalization with the last year	238	44	1.0	–	17	1.0	–
Infectious disease	144	60	2.7*	1.2-5.8	33	2.1*	1.2-3.8
Noninfectious disease	77	25	3.9**	2.1-7.3	13	3.2**	2.0-5.0

Table 1 (continued)
Exposure: E (+)	Total E (+)	CZ-RES			CTX-RES
		Cases	OR	95% CI	Cases	OR	95% CI
Prior infectious sites
None	316	69	1.0	–	30	1.0	–
Low respiratory tract	20	8	2.4	0.9-6.1	6	4.1**	1.5-11.4
Intra-abdominal	45	12	1.3	0.6-2.7	7	1.8	0.7-4.3
Urinary tract	34	12	2.0	0.9-4.1	8	2.9*	1.2-7.1
Others	26	16	5.7**	2.5-13.2	9	5.1**	2.1-12.3
Prior antibiotic exposure	166	69	2.8**	1.8-4.2	37	3.0**	1.7-5.1
Prior infection with resistant strain
CZ-RES	25	12	2.5*	1.1-5.7	7	2.6*	1.1-6.6
CTX-RES	56	28	3.0**	1.7-5.3	17	3.4**	1.8-6.5
Vancomycin needed at OPD	20	13	5.2**	2.0-13.3	8	4.7**	1.8-11.9
Procedure at OPD
No procedure needed at OPD	423	115	1.0	–	56	1.0	–
Chemotherapy at OPD	19	3	0.5	0.1-1.8	1	0.4	0.1-2.8
Hemodialysis	7	4	1.9	0.9-4.0	2	1.6	0.7-3.7
Others	11	7	0.6	0.1-5.8	4	1.6*	1.0-2.4
Mechanical valve	5	1	0.6	0.1-5.7	0	–	–
Arteriovenous fistula for dialysis	9	4	2.1	0.6-8.0	2	1.8	0.4-9.0
Intravascular catheter	7	5	3.1**	1.7-5.6	3	3.0**	1.5-5.9
Foley catheter	16	10	4.6**	1.6-12.8	8	7.1**	2.6-19.6
Other urinary drainage	5	3	3.9	0.7-23.7	0	–	–
Missed discharged at ED*	34	10	1.1	0.5-2.3	5	1.1	0.4-2.9
Recalling to ED	20	8	1.8	0.7-4.4	4	1.6	0.5-5.0
Initial presentations
Fever duration
No fever	124	36	1.0	–	20	1.0	–
0-3 d	294	77	0.9	0.5-1.4	35	0.7	0.4-1.3
N4 d	42	16	1.5	0.7-3.1	8	1.2	0.5-3.0
Systolic blood pressure (mm Hg)
Low (V90)	96	39	1.9*	1.2-3.2	21	1.8	1.0-3.4
Normal (91-140)	209	55	1.0	–	28	1.0	–
High (z141)	155	35	0.8	0.5-1.3	14	0.6	0.3-1.3
Diastolic blood pressure (mm Hg)
Low (V60)	72	38	3.4**	2.0-5.7	20	2.9**	1.6-5.3
Normal (60-89)	330	82	1.0	–	39	1.0	–
High (z90)	58	9	0.6*	0.3-1.2	4	0.6	0.2-1.6
Body temperature (8C)
b35.9	31	14	2.2*	1.0-4.7	9	2.6*	1.1-6.1
36-38.2	210	58	1.0	–	29	1.0	–
38.3-39.4	154	43	1.0	0.6-1.6	20	0.9	0.5-1.7
N39.5	65	14	0.7	0.4-1.4	5	0.5	0.2-1.4
Heart rate (per min)
V59	5	1	0.6	0.1-5.8	1	1.8	0.2-17.2
60-99	140	40	1.0	–	17	1.0	–
z100	315	88	1.0	0.6-1.5	45	1.0	0.7-2.2
Oxygen saturation
Spo2 = 96-100%	426	118	1.0	–	57	1.0	–
Spo2 b 95%	34	11	1.5	1.0-2.3	6	2.4**	1.4-4.1
WBC count (109/L)
V1000	20	11	4.0**	1.6-10.1	8	6.7**	2.5-17.9
1001-5000	47	13	1.3	0.6-2.5	10	2.7**	1.2-6.1
5001-15000	287	67	1.0	–	26	1.0	–
z15001	106	38	1.8*	1.1-3.0	19	2.2*	1.2-4.2
Neutrophil percentage of WBC
(continued on next page)

Table 1 (continued)
Exposure: E (+)	Total E (+)	CZ-RES				CTX-RES
		Cases	OR	95% CI		Cases	OR	95% CI
V49	25	17	6.3**	2.3-16.9		12	6.4**	2.2-18.2
50-80	71	18	1.0	–		9	1.0	–
z81	355	90	1.0	0.6-2.0		40	0.9	0.4-1.9
Band-from percentage of WBC
No band-form cell	337	94	1.0	–	48		1.0	–
119%	31	7	0.8	0.3-1.8	1		0.1	0.0-1.5
z10%	83	23	1.0	0.6-1.7	12		1.0	0.5-2.0
Hemoglobin level (g/dL)
b10	146	51	1.6*	1.0-2.5	32		2.6**	1.5-4.5
10.1-12.9	272	68	1.0	–	27		1.0	–
N13	42	10	0.9	0.4-2.0	4		1.0	0.3-2.9
Thrombocyte count (109/L)
V99	123	40	1.4	0.9-2.2	22		1.6	0.9-2.9
100-399	320	82	1.0	–	38		1.0	–
z400	17	7	2.0	0.8-5.5	3		1.6	0.4-5.8

ical variables and the outcomes (CZ-RES and CTX-RES) were then assessed using the m² test or Fisher exact test, as appropriate. If a test was significant ( P b .05), the corresponding variable was considered to be univariately predictive of outcome. All variables showing an association with outcome ( P b .25) were entered into the multivariate logistic regression analysis.

OPD indicates outpatient department; WBC, white blood cell.

* P b .05

** P b .01

We used stepwise method to select the parsimonious model. Collinearity of covariates was assessed by the variance inflation factor. receiver operating characteristic curves were constructed for assessing Predictive validity on the basis of the area under the curve (AUC) and their confidence intervals (CIs) calculated with Hanley- McNeil formula [26]. The model fitting was assessed using the Hosmer-Lemeshow goodness-of-Fit test.

To make the prediction models easy to use, we simplified them to make a prediction rule [27,28] using a regression coefficient-based scoring method. With this, a simple integer-based point score was generated for each predictor by doubling the value of the corresponding b-coefficient in

the model and rounding up to the nearest integer. The overall risk score was calculated by adding the individual scores for each of the predictors together [29]. The resulting total scores were grouped to demonstrate the Predictive performance of the models. All analyses were performed using SAS software version 9.1 (SAS Institute Inc, Cary, NC).

Results

There was a total of 1346 episodes of positive bacteremia identified at the ED during the study period, of which 695 episodes were of Gram-negative infection from 659 patients, with 25 patients having repeated episodes within a year. As shown in Fig. 2, these 695 episodes were randomly allocated (in the ratio of 2:1) to the derivation and validation groups. Univariate analyses of collected variables were listed in Table 1. Of the demographic characteristics, there were 349 men and 310 women; average age of study population was 63.9 F

Table 2 Predictors and assigned scores from the multivariate analysis of CZ-RES model
Parameter	b-Coefficient	OR	95% CI	P	Scores
Time since last discharge: within 7 d	2.1492	8.6	4.2-17.4	b.0001	4
Time since last discharge: between 8 and 30 da	1.5444	4.7	2.4-9.0	b.0001	3
Immunosuppressant use	1.3610	3.9	1.0-15.1	.0488	3
Prior infection with a CTX-RES strain	0.9694	2.6	1.4-5.0	.0039	3
Residence in a nursing home or stroke with history of repeated choking	1.1062	3.0	1.6-5.7	.0006	2
Oxygen saturation b95%a	0.6009	1.8	1.1-3.0	.0209	1
Liver cirrhosis	–1.2143	0.2	0.1-0.6	.0043	–2
a Nonpredictors or different assigned scores in comparing to CTX-RES model.

Table 3 Predictors and assigned scores for the CTX-RES model
Parameter	b-Coefficient	OR	95% CI	P	Scores
Time since last discharge: within 7 d	2.2256	9.3	4.3-20.1	b.0001	4
Time since last discharge: between 8 and 14 d*	1.1744	3.2	0.9-11.2	.063	2
Prior infection with a CTX-RES strain	1.2874	3.6	1.6-8.1	.0017	3
Immunosuppressant use	1.6440	5.2	0.9-28.8	.0606	3
WBC count N1000/mm3*	1.3278	3.9	1.8-8.8	.0008	3
Oxygen saturation b95%*	1.2155	3.4	1.7-6.6	.0004	2*
Residence in a nursing home or stroke with history of repeated choking	1.0725	2.9	1.3-6.4	.0078	2
WBC count N15000/mm3*	0.7170	2.0	1.0-4.2	.0515	1
Liver cirrhosis	–1.2336	0.2	0.1-1.1	.0671	–2
* Nonpredictors or different assigned scores in comparing to CZ-RES model.

16.9 years (mean F SD). In the univariate analysis, men were more likely to develop CZ-RES (odds Ratio [OR], 1.5; 95% CI, 1.1-2.3) compared to women, but this was not statistically significant in the multivariate regression. Some underlying comorbidities also showed a strong association with resistance, in particular, having a history of stroke with repeated choking or hematological malignancy. Cirrhosis of liver showed a significant inverse association with CZ-RES but not with CTX-RES in univariate analysis. Many of the hospital exposure and health care-related factors were significantly associated with elevated risk of developing resistance, the strongest indicators being time since last hospitalization and previous infection with a resistant strain. Of the initial presentations, we found that Abnormal vital signs, and abnormally high or low complete blood cell and differential counts were associated with greater risk. Poor oxygen saturation was related to CTX- RES but not to CZ-RES.

Final predictors and assigned coefficient-based scores for CZ-RES model and CTX-RES model were listed in Table 2 and Table 3. The covariates for CZ-RES model were discharge from hospital within the last month, prior

infection with a CTX-RES strain, residence in a nursing home or history of stroke with repeated choking, use of immunosuppressants in transplantation patients, leucopenia [b1000 (1/mm)] or leukocytosis [N15000 (1/mm)] at first blood test, and poor oxygen saturation (Spo₂ b95%). Surprisingly, after adjusting for other factors in the multivariate analysis, cirrhosis of the liver had a protective effect with regard to CZ-RES and was retained in the final model. For ease of clinical use, it was also kept in the final CTX-RES model, although it was only marginally signif- icant there. Predictors included in the final CTX-RES model were all those already included in the CZ-RES model plus abnormal leukocyte count (b1000 or N15000/ mm³) at first blood test. Besides, shorter interval since prior hospitalization, particularly within 2 weeks, led to an increase risk for antimicrobial-resistant bacteremia. Hosmer-Lemeshow goodness-of-fit test shows a satisfacto- ry model fitting ( P = .83 for CZ-RES and P = .28 for CTX-RES).

The ROC curves for the CZ-RES and CTX-RES models

were plotted in Fig. 3 and Fig. 4, respectively. For the CZ- RES models, the AUC of the primary coefficient model

Fig. 3 The ROC curves for assessing predictive validity of CZ- RES models.

Fig. 4 The ROC curves for assessing predictive validity of CTX- RES models.

Range of scores		–2 to 1	2-3	4-5	6-7	z8	Total
Derivation	CZ-RES	43	33	29	22	2	129
	Subtotal	290	90	49	28	3	460
	Risk	15%	37%	59%	79%	67%	28%
Validation	CZ-RES	24	25	10	6	2	67
	Subtotal	155	46	21	10	3	235
	Risk	15%	54%	48%	60%	67%	29%

(abbreviating to coefficient derivation) was 0.76 (95% CI, 0.71-0.81). For the CTX-RES models, the AUC of the primary coefficient model (coefficient derivation) was 0.82 (95% CI, 0.76-0.88). The closeness of the 4 ROC curves (coefficient and score models applied to both the derivation and validation data sets) in both models suggests good predictive validity.

Table 4 CZ-RES scoring model: predictive performance for different scores

To develop the scoring models into a practical tool for clinical decision making, the optimal cut points were identified, and the predictive accuracy at these points was assessed. For the CZ-RES model, from the derivation data, we estimated this to be bscore z2Q (giving sensitivity and specificity of 67% and 75%, respectively), and from the validation set, bscore z3Q (reducing sensitivity to 60% but increasing specificity to 80%). Applying the same approach to the CTX-RES model, we found the equivalent optimal cut points to be score z3 (yielding 81% sensitivity and 65% specificity) and bscore z4,Q which yielded 63% sensitivity and 84% specificity.

To assess whether the performance of this Decision tool is adequate for clinical use, we ascertained its Predictive ability for a range of different scores. As shown in Tables 4 and 5, the overall proportion in the derivation and validation

sets with CZ-RES GNB was 28% and 29%, respectively. However, the results by score indicate that the proportion with resistant infection increases with score and that the decision rules, therefore, has good predictive properties. There was a slight decrease in the proportion with resistant infection in the bscore z8Q group, but this may well be due to the fact that so few were in this category. Similarly, the proportion with CTX-RES GNB in the derivation and validation data sets was 14% and 17% respectively. Again, the score z8 group was relatively small, but the resistant rate otherwise increased with score. Thus, it appears that the decision rules we developed do have sufficient Predictive power to make them useful in clinical setting, that is, they are able to differentiate between the cases most likely to have resistant or nonresistant disease.

Discussion

In this study, we used a derivation-validation method to develop 2 decisions rules to predict antibiotic-resistant GNB in patients at ED by a simple scoring system that uses only the clinical information that is readily available

Range of scores		2-3	2-3	4-5	6-7	z8	Total
Derivation	CTX-RES	7	16	15	12	13	63
	Subtotal	200	157	58	29	16	460
	Risk	4%	10%	26%	41%	81%	14%
Validation	CTX-RES	7	8	13	7	4	39
	Subtotal	120	61	33	11	10	235
	Risk	6%	13%	39%	64%	40%	17%

at the time. We tried to adhere to the standard criteria in development of prediction model [26,27]. To the best of our knowledge, it is also the first study to elucidate the association between clinical risk factors and antimicrobial resistance in Taiwan. Our prediction models appear to be useful for clinical practice by incorporating relevant risk factors.

Table 5 CTX-RES scoring model: predictive performance for different scores

Because GNB may ensue from inadequate antibiotic treatment, the choice of first-, second-, or third-generation antibiotic is of paramount importance, yet the decision often has to be made before the culture result is available. Physicians tend to prescribe a second generation of antibiotic for infected patients with risk factors suggestive of resistance. However, if the patient is stable clinically and the doctor is unaware of a risk factor (eg, recent hospitalization within the last month), he may be in a dilemma as to which antibiotic to prescribe. Our CZ-RES and CTX-RES models, which were derived from a large study according to respected scientific guidelines, provide a good evidence-based reference in this circumstance. Such a model would enable physicians to commence treatment earlier and therefore reduce morbidity, mortality, and the related economic costs.

Hospital exposure and health care-associated factors, as widely documented in previous literature, were also significant in our cohort. Friedman et al [11] found that patients living in nursing home had similar Bloodstream infections to those in hospital settings, particularly meth- icillin-resistant Staphylococcus aureus. However, the anti- biotic resistance of Gram-negative pathogens was not further analyzed in their study. From a case-control study, Schippa et al [13] reported an OR of 3.6 for ceftazidime- resistant GNB in nursing home residents, comparing to general population. In our prospective cohort, although nursing home residence was significant in univariate analysis, it lost significance after being adjusted by other covariates (OR, 1.6; 95% CI, 0.6-6.2). Similar results were found for history of stroke with repeated choking, but after dealing with the problem of collinearity of these 2 variables by selecting a composite measure, the factor turned out to be significant in final models.

Associations between initial clinical presentation and antimicrobial resistance have barely been addressed before. Leukocytopenia (b1000/mm³) has long been suspected of being related to serious Gram-negative infection (such as Pseudomonas spp). Besides, Jaimes et al [30] found

leukocytosis (N12 000/mm³) to be an independent predictor of positive bacteremia but did not investigate the effect on antibiotic resistance. We found that leucopenia (b1000/ mm³) and leukocytosis (N15000/mm³) were both associated with antimicrobial-resistant GNB. Theoretically, bacteremia leads to a more severe systemic response than focal infections, but whether cephalosporin resistant pathogens have the same effect on each case is still under investigation. Poor oxygen saturation was also related to antimicrobial resistance in our study, probably because of its association with severe pneumonia. However, because the infectious focus might be unknown initially at ED, we did not take this variable (ie, final diagnosis) into account when developing our prediction rules.

Liver cirrhosis was inversely associated with cephalo- sporin-resistant GNB in our study, although many studies have emphasized that their outcomes are poorer in comparison to the general population. The poor outcome is perhaps due to the Severity of disease rather than the infectious pathogen [31-34]. Graudal et al reported that there was no difference of bacterial strains when bacteremia occurred in Cirrhotic patients versus noncirrhotic patients [35]. Because of chronic hepatitis B viral infection, liver cirrhosis is relatively common in the Chinese and Taiwanese. Our study includes a larger proportion of cirrhotic patients than would be found in western studies, which may partly explain the novel finding of inverse association. Of course, the biologic rationale should be clarified in future research.

There are several limitations to this study. Firstly, there is a lack of clinical data from patients with negative blood cultures at ED. Thus, the decision rules developed here are not applicable to all ED patients with suspected blood infection because the ED deals with more than 100000 patients per year and our research team was unable to follow-up those whose blood cultures turned out to be negative. Secondly, we did not collect information on the antimicrobial agents prescribed by the primary care physi- cians at ED. If this data were available, we would have been able to compare the accuracy of our model predictions with that of Clinician judgment. Thirdly, the questionnaire used to collect medical history data was constructed by the principal researchers specifically for this study, so it was never formally validated. Finally, because our samples were derived from the ED of a tertiary hospital, our results may not be representative of patients in other community hospitals. Antimicrobial resistance is an extremely complex problem worldwide. Multicenter studies may be needed in the future.

Conclusion

We have developed 2 models for predicting risk of antimicrobial Gram-negative infection by identifying and quantifying associated risk factors. These models could be

used by physicians to determine the most appropriate choice of antibiotic for first-line therapy, particularly in situations where the culture result is not yet known.

References

1. Carmeli Y, Troillet N, Karchmer AW, et al. Health and economic outcomes of antibiotic resistance in Pseudomonas aeruginosa. Arch Intern Med 1999;159:1127 – 32.
2. Lautenbach E, Patel JB, Bilker WB, et al. Extended-spectrum beta- lactamase-producing Escherichia coli and Klebsiella pneumoniae: risk factors for infection and impact of resistance on outcome. Clin Infect Dis 2001;32:1162 – 71.
3. Virk A, Steckelberg JM. Clinical aspects of antimicrobial resistance. Mayo Clin Proc 2000;75:200 – 14.
4. Zuccotti GV, Gracchi V, Sala D, et al. Nosocomial infections at a pediatric age. Pediatr Med Chir 2002;24:177 – 85.
5. Weinstein MP, Towns ML, Quartey SM, et al. The clinical significance of positive blood cultures in the 1990s: a prospective comprehensive evaluation of the microbiology, epidemiology, and outcome of bacteremia and fungemia in adults. Clin Infect Dis 1997;21:584 – 602.
6. Pop-Vicas AE, D’Agata EM. The rising influx of multidrug-resistant Gram-negative bacilli into a tertiary care hospital. Clin Infect Dis 2005;40:1792 – 8.
7. Lucet JC, Chevert S, Decre D, et al. Outbreak of multiply resistant Enterobacteriaceae in an intensive care unit: epidemiology and risk factors for acquisition. Clin Infect Dis 1996;22:430 – 6.
8. De Champs C, Rouby D, Guelon D, et al. A case-control study of an outbreak of infections caused by Klebsiella pneumoniae strains producing CTX-1(TEM-3) beta-lactamase. J Hosp Infect 1991; 18:5 – 13.
9. Paterson DL, Mulazimoglu L, Casellas JM, et al. Epidemiology of ciprofloxacin resistance and its relationship to extended-spectrum beta-lactamase production in Klebsiella pneumoniae isolates causing bacteremia. Clin Infect Dis 2000;30:437 – 8.
10. Wiener J, Quinn JP, Bradford PA, et al. Multiple antibiotic-resistant Klebsiella pneumoniae and Escherichia coli in nursing home. JAMA 1999;281:517 – 23.
11. Friedman ND, Kaye KS, Stout JE, et al. Health care-associated bloodstream infections in adults: a reason to change the accepted definition of Community-acquired infections. Ann Intern Med 2002;137:791 – 7.
12. Siegman-Igra Y, Fourer B, Orni-Wasserlauf R, et al. Reappraisal of community-acquired bacteremia: a proposal of a new classification for the spectrum of acquisition of bacteremia. Clin Infect Dis 2002; 34:1431 – 9.
13. Schiappa DA, Hayden MK, Matushek MG, et al. Ceftazidime- resistant Klebsiella pneumoniae and Escherichia coli bloodstream infection: a case-control and molecular epidemiologic investigation. J Infect Dis 1996;174:529 – 36.
14. Blot S, Vandewoude K, De Bacquer D, et al. Nosocomial bacteremia caused by antibiotic-resistant Gram-negative bacteria in critically ill patients: clinical outcome and length of hospitalization. Clin Infect Dis 2002;34:1600 – 6.
15. Pena C, Pujol M, Ricart A, et al. Risk factors for faecal carriage of Klebsiella pneumoniae producing extended spectrum beta-lactamase (ESBL-KP) in intensive care unit. J Hosp Infect 1997;35:9 – 16.
16. Piroth L, Aube H, Doise JM, et al. Spread of extended-spectrum beta-lactamase-producing Klebsiella pneumoniae : are beta- lactamase inhibitors of therapeutic value? Clin Infect Dis 1998; 27:76 – 80.
17. Ho M, Hsiung CA, Yu HT, et al. Antimicrobial usage in ambulatory patients with Respiratory infections in Taiwan, 2001. J Formos Med Assoc 2004;103:96 – 103.
18. Goldberg D. The cephalosporins. Med Clin North Am 1987; 71:1113.
19. Gantz NM, Brown R, Berk SL, Esposito AL, et al. Manual of clinical problems in infectious disease. Chap. 74: choosing a cephalosporin, 4th ed. 1999. p. 410 – 4.
20. MacDonald KL, Cohen ML, Hargrett-Bean NT, et al. Changes in antimicrobial resistance of Salmonella isolated from humans in the United States. JAMA 1987;11:1496 – 9.
21. Ackers ML, Puhr ND, Tauxe RV, et al. Laboratory-based surveillance of Salmonella serotype Typhi infections in the United States: antimicrobial resistance on the rise. JAMA 2000;283:2668 – 73.
22. Ko WC, Lee HC, Chuang YC, et al. Clinical features and Therapeutic implications of 104 episodes of monomicrobial Aeromonas bacter- aemia. J Infect 2000;40:267 – 73.
23. Weber JT, Mintz ED, Canizares R, et al. Epidemic cholera in Ecuador: multidrug-resistance and transmission by water and seafood. Epide- miol Infect 1994;112:1 – 11.
24. van den Bogaard AE, Stobberingh EE. Antibiotic usage in animals: im- pact on bacterial resistance and public health. Drugs 1999;58:589 – 607.
25. Sharma H, Unicomb L, Forbes W, et al. Antibiotic resistance in Campylobacter jejuni isolated from humans in the Hunter Region, New South Wales. Commun Dis Intell 2003;27(Suppl):S80-8.
26. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143:29 – 36.
27. Wasson JH, Sox HC, Neff RK, et al. clinical prediction rules. Applications and methodological standards. N Engl J Med 1985;313:793 – 9.
28. Laupacis A, Sekar N, Stiell IG. Clinical prediction rules. A review and suggested modifications of methodological standards. JAMA 1997;277:488 – 94.
29. Sullivan LM, Massaro JM, D’Agostino Sr RB. Presentation of multivariate data for clinical use: The Framingham Study risk score functions. Stat Med 2004;23:1631 – 60.
30. Jaimes F, Arongo C, Ruiz G, et al. Predicting bacteremia at the bedside. Clin Infect Dis 2004;38:357 – 62.
31. Aloy Duch A, Espejo Arenas E, Mauri Pont M, et al. Bacteremia in the patient with liver cirrhosis. Prospective study of 61 episodes. Enferm Infecc Microbiol Clin 1990;8:540 – 3.
32. Kuo CH, Changchien CS, Yang CY, et al. Bacteremia in patients with cirrhosis of the liver. Liver 1991;11:334 – 9.
33. Mizuno R, Nishiyama Y, Shimizu S, et al. Clinical study on bacteremia in patients with liver cirrhosis. Kansenshogaku Zasshi 1996;70:456 – 62.
34. Thulstrup AM, Sorensen HT, Schonheyder HC, et al. Population-based study of the risk and short-term prognosis for bacteremia in patients with liver cirrhosis. Clin Infect Dis 2000;31:1357 – 61.
35. Graudal N, Milman N, Kirkegaard E, et al. Bacteremia in cirrhosis of the liver. Liver 1986;6:297 – 301.