Risk factors for development of aminoglycoside resistance among gram-negative rods

Abstract

Purpose

Development of scoring systems to predict the risk of aminoglycoside resistance and to guide therapy is described.

Methods

Infections due to aminoglycoside-resistant gram-negative rods (AR-GNRs) are increasingly common and associated with adverse outcomes; selection of effective initial antibiotic therapy is necessary to reduce adverse consequences and shorten length of stay. To determine risk factors for AR-GNR recovery from culture, cases of GNR infection among patients admitted to 2 institutions in a major academic hospital system during the period 2011–2016 were retrospectively analyzed. Gentamicin and tobramycin resistance (GTR-GNR) and amikacin resistance (AmR-GNR) patterns were analyzed separately. A total of 26,154 GNR isolates from 12,516 patients were analyzed, 6,699 of which were GTR, and 2,467 of which were AmR.

Results

In multivariate analysis, risk factors for GTR-GNR were presence of weight loss, admission from another medical or long-term care facility, a hemoglobin level of <11 g/dL, receipt of any carbapenem in the prior 30 days, and receipt of any fluoroquinolone in the prior 30 days (C statistic, 0.63). Risk factors for AmR-GNR were diagnosis of cystic fibrosis, male gender, admission from another medical or long-term care facility, ventilation at any point prior to culture during the index hospitalization, receipt of any carbapenem in the prior 30 days, and receipt of any anti-MRSA agent in the prior 30 days (C statistic, 0.74). Multinomial and ordinal models demonstrated that the risk factors for the 2 resistance patterns differed significantly.

Conclusion

A scoring system derived from the developed risk prediction models can be applied by providers to guide empirical antimicrobial therapy for treatment of GNR infections.

KEY POINTS

• 26,154 gram-negative rod isolates from 12,516 patients were analyzed to determine risk factors for resistance to gentamicin and tobramycin and resistance to amikacin.

• Resistance risk factors included male gender, cystic fibrosis, weight loss, admission from another medical facility, anemia, and recent receipt of carbapenems, fluoroquinolones, or anti-MRSA agents.

• Multinomial and ordinal models demonstrated that the factors predicting resistance to gentamicin and tobramycin and resistance to amikacin were significantly different.

The increasing global rate of human infections with multidrug-resistant organisms (MDROs) is associated with increased adverse outcomes and cost.¹ In the United States alone, there is an estimated $50 million in yearly costs and 23,000 yearly deaths attributable to MDRO infections.² Initial use of appropriate antibiotics decreases mortality and hospital length of stay,^3,4 while overuse of broad-spectrum antibiotics has been linked with increased prevalence of MDROs^5–9; the initial antibiotic choice for empirical coverage remains a challenging and high-stakes decision.

Aminoglycosides are a class of antibiotics typically reserved for treatment of gram-negative isolates resistant to β-lactams and other first-line antibiotic classes.¹⁰ Aminoglycoside resistance among gram-negative rods (AR-GNRs) has increased over the past several decades^11–13 and typically co-occurs with resistance to other antibiotics.^14,15 This resistance is associated with an increased risk of inappropriate initial antibiotic therapy, which has been shown to increase length of stay and mortality.^16–19 Since aminoglycoside resistance patterns are not available immediately and risk factors for AR-GNR infection have been incompletely studied, it can be difficult to know when to account for the possibility of AR-GNR infections.

Prior literature on risk factors for aminoglycoside-resistant infections has been sparse. Several risk factors have been identified, primarily exposure to aminoglycosides,^14,20–22 higher-level care,²⁰ and the presence of indwelling devices and exposure to other antibiotics.^15,23 Other risk factors for development of MDROs in general include prior residence in a nursing home, hemodialysis, intensive care unit (ICU) admission,²⁴ an increased number of medical comorbidities,^25,26 prior antibiotic usage, and invasive surgery.²⁶

The primary purpose of the study described here was to examine risk factors for resistance and create models and scoring systems to predict resistance to aminoglycosides with the goal of improving the effectiveness of initial empirical antibiotic therapy. Our secondary purpose was to compare and contrast the risk factors associated with resistance to different aminoglycosides to determine if all aminoglycoside resistance is driven by similar risk factors.

Methods

Population and data

In order to develop a comprehensive model for risk of infection or colonization with AR-GNRs, we conducted a retrospective study of all patients at 2 hospitals in metropolitan Los Angeles, California with positive cultures from any source over a nearly 6-year period. We performed 2 separate analyses: one for gentamicin- and tobramycin-resistant gram-negative rods (GTR-GNRs) and the other for amikacin-resistant gram-negative rods (AmR-GNRs). Ronald Reagan UCLA Medical Center is a 520-bed tertiary care center that has 5 adult ICUs with a total of 109 beds. Santa Monica UCLA Medical Center has 266 beds in total, with 22 mixed intensive care beds in a single unit. Both hospitals are in the UCLA Health system and serve patients with various medical and surgical conditions, including solid organ and bone marrow transplant recipients and patients with cancer. The UCLA Health Integrated Clinical and Research Data Repository serves as a warehouse for all clinical data in the system’s electronic medical record since 2006. Only data from 2011 and onwards were used in the study, as a changeover in clinical data warehousing methods that year resulted in significantly more robust clinical information. The initial data set contained information from all admissions to either hospital of patients 18 years of age or older and at least 1 positive culture from any source (blood, urine, sputum, wound cultures, or other fluids) with start dates from January 2011 through November 2016.

Our sample included all applicable cases within the period during which we had access to complete data and, as such, represents the complete universe of cases for our institutions. While sample size calculations for multivariate logistic regression are not straightforward, a validated rule recommends at least 10 cases per predictor variable.²⁷ Given our sample size of more than 10,000 for each analysis, we felt predictive power was more than adequate for the models we sought to build.

Since the endpoint of the analysis was prediction of development of the first aminoglycoside-nonsusceptible isolate, once a patient had a culture growing an organism with nonsusceptibility to the antibiotic in question, as defined using Clinical and Laboratory Standards Institute (CLSI) breakpoints current in the year of testing, all cultures from that patient occurring at a later time than the original culture were removed from the data set. The concordance of gentamicin and tobramycin resistance in our data set was 94.3%. Since gentamicin and tobramycin had significant overlap in resistance patterns and neither drug was consistently more effective than the other, gentamycin resistance and tobramycin resistance were treated as a single entity, with an organism categorized as GTR if it had resistance to either gentamicin or tobramycin.

Routine susceptibility testing was performed by the CLSI reference broth microdilution method using panels prepared in-house. All antimicrobial susceptibility data were interpreted using CLSI breakpoints current to the year of testing. Isolates with intrinsic nonsusceptibility to the studied antibiotics were included in the analysis and coded as nonsusceptible.

The list of examined predictor variables was drawn from 2 sources: significant predictors from prior similar studies and predictors with biologic plausibility that were readily obtained from the medical record at the study institutions. Data collected for each patient included admission hospital, days since admission, location prior to admission (home versus long-term care facility or other hospital), demographic information, comorbidities (grouped into categories by Elixhauser comorbidity score designations),²⁸ laboratory results from the date of the culture, vital signs on the date the culture was collected (maximum temperature, heart rate, respiratory rate, and minimum blood pressure), oxygenation and/or ventilation method, presence of indwelling devices (tracheostomy or urinary catheter), administration of antibiotics and other selected medications (vasopressors, probiotics, blood products, immunosuppressants, and acid suppressants), and culture source.

An infection was coded as “hospital-acquired” if the culture was submitted to the laboratory more than 48 hours after the time of the patients’ first presentation to the hospital. The variable “advanced ventilatory support” referred to patients receiving either noninvasive or invasive mechanical ventilation. Administration of various medications (including antibiotics) was coded as the number of days since last receipt of the medication, with values Winsorized to a maximum value of 100 (0 denoted receipt within 24 hours of the time of culture; nonreceipt of the medications of interest was coded as 100 days since receipt). “Anti-MRSA” agents referred to i.v. vancomycin, linezolid, and daptomycin, as these were used at the study institutions in cases of suspected hospital-acquired methicillin-resistant Staphylococcus Aureus (MRSA) infection. These antibiotics were grouped by function to improve predictive power, since they were at times used interchangeably at the institutions in particular patient populations (such as those with solid organ or bone marrow transplants).

Antibiotics were considered to have been administered if they were received by any route, including the inhaled, i.v., and oral routes.

In cases where laboratory tests were not performed before cultures were sent (typically on the first day of a patient’s admission), the first set of laboratory results was used, provided that tests were performed on specimens collected within 24 hours of obtaining a positive culture. For laboratory tests not typically performed daily (e.g., liver function tests, measures of coagulation, and protein or prealbumin determinations), the most recent result within a 48-hour period prior to obtaining a positive culture was used.

To facilitate model interpretability, linear variables were recategorized as binary: either above or below a cutoff threshold. Various cutoffs were tested against one another in the final model (e.g., 30 versus 60 versus 90 days since receipt of the last antibiotic), and the cutoff value that led to the highest C statistic was chosen for inclusion in the scoring system.

Statistical analysis

Compared to gentamicin and tobramycin, amikacin is typically associated with higher rates of nonsusceptibility.^11,13,29 It is unclear from prior literature if the risk factors for GTR are the same as for AmR. Therefore, 2 separate analyses were performed, one examining risk factors for GTR-GNR and one examining risk factors for AmR-GNR, to examine the similarities and differences in risk factors for recovery of a nonsusceptible isolate. These analyses were followed by an analysis to determine if the risk factors were similar enough that a single model could predict both types of resistance.

Two separate analyses were performed, one comparing all gentamicin- and tobramycin-sensitive GNR (GTS-GNR) against GTR-GNR, and one comparing amikacin-susceptible GNR (AmS-GNR) against AmR-GNR. These analyses were chosen to aid decisions at the point of initial antibiotic choice, when the consequences for inappropriate antibiotic therapy are the greatest.^16–19 Such decisions must be undertaken before any information is available regarding the species and resistance patterns of the causative agent. Therefore, organisms with various resistance mechanisms and organisms with intrinsic resistance (e.g., Stenotrophomonas species) were all included in the analysis, since there is no way of clinically differentiating between them prior to making the initial antibiotic choice.

The measured variables in each case were compared between the cases and controls by a 2-sided Mann-Whitney U test, Student’s t test, or a chi-square test, as appropriate. In each case, after bivariate associations were examined, variables with p < 0.10 or strong biologic plausibility were included in a stepwise forward model selection procedure to create a logistic regression model for each analysis. A p value of <0.10 was chosen as a threshold for inclusion of variables in the model to ensure that marginally significant variables with potentially good predictive power were included. Only complete cases were included in model selection. Model discrimination was assessed via area under the receiver operating characteristic curve (C statistic) analysis, and models were compared by chi-square test if they were nested or by Akaike information criterion (AIC) if they were not. For all other statistical tests, an a priori p value of <0.05 was used.

For the combined model, 3 categories were created: susceptible to all aminoglycosides, GTR but not AmR, and AmR. If an isolate was nonsusceptible to amikacin but susceptible to gentamicin and tobramycin, it was classified as AmR as though it were also sensitive to gentamicin and tobramycin; these isolates comprised less than 3% of the AmR isolates in total. This decision was made to aid in model interpretability given that ordinal logistic models require a clear hierarchy of outcomes and that organisms expressing AmR would likely not be treated with any aminoglycoside as long as alternative therapies were available.

Multinomial and ordinal logistic regression models were then fitted for prediction of all 3 outcome categories, and these models were tested against each other by AIC. The Brant test was used to determine if the proportional odds assumption held for the ordinal model. All analyses were performed using the Stata statistical software package, version 14.2 (StataCorp LP, College Station, TX).

Results

Results overview

The complete data set included 26,154 GNR isolates from 12,516 patients, 6,699 (25.6%) of which were GTR, and 2,467 (9.4%) of which were AmR. After removing redundant cultures from the GTR analysis, there were 25,627 isolates remaining, of which 4,087 (15.9%) were GTR-GNR. After removing redundant cultures from the AmR analysis, there were 25,002 isolates remaining, of which 1,315 (5.3%) were AmR-GNR. There were substantially lower percentages of resistant isolates in the final models because all cultures after the first resistant isolate for a given model were dropped. Additionally, only complete cases were analyzed for the multivariate models. The final model for GTR comprised 12,457 cultures, 1,779 (14.3%) of which were GTR-GNR, and the final model for AmR comprised 12,062 cultures, 532 (4.4%) of which were AmR-GNR. The multinomial model comprised 12,062 cultures, 1,973 (16.4%) of which were GTR-GNR but not AmR-GNR and 532 (4.4%) of which were AmR-GNR.

The most common GTR-GNR isolates were Escherichia species, while the most common GTS-GNR isolates were Escherichia and Pseudomonas species; the most common AmR-GNR isolates were Stenotrophomonas species (which have intrinsic resistance), while the most common AmS-GNR isolates were Escherichia species (Table 1). Respiratory culture source was predictive of both the GTR-GNR and AmR-GNR classifications, while urinary source was predictive of both the GTS-GNR and AmS-GNR classifications (Table 2).

Table 1.

Distribution of Isolates by Resistance and Susceptibility Classification^a,b

Genus	GTS (n = 21,540)	GTRc (n = 4,087)	AmS (n = 23,687)	AmRc (n = 1,315)
Acinetobacter	212 (1.0)	302 (7.4)	240 (1.0)	230 (17.5)
Enterobacter	2,052 (9.5)	55 (1.3)	2,128 (9.0)	5 (0.4)
Escherichia	7,510 (34.9)	1,411 (34.5)	9,633 (40.7)	76 (5.8)
Klebsiella	3,226 (17.8)	509 (12.5)	4,352 (18.4)	195 (14.8)
Pseudomonas	4,669 (21.7)	447 (10.9)	3,284 (13.9)	126 (9.6)
Stenotrophomonas	0	568 (13.9)	0	541 (41.4)
Other	3,871 (15.1)	795 (19.4)	4,050 (17.0)	142 (10.8)

^aGTS = gentamycin and tobramycin sensitive, GTR = gentamycin and tobramycin resistant, AmS = amikacin sensitive, AmR = amikacin resistant.

^bAll data are number and percentage.

^c p < 0.001 for χ ² test.

Table 2.

Distribution of Culture Sources by Resistance and Susceptibility Classification of Isolates^a,b

Source	GTS-GNR (n = 21,540)	GTR-GNRc (n = 4,087)	AmS-GNR (n = 23,687)	AmR-GNRc (n = 1,315)
Blood	1,893 (8.8)	319 (7.8)	2,111 (8.9)	89 (6.8)
Urine	9,947 (46.2)	1,580 (38.7)	11,468 (48.4)	195 (14.8)
Respiratory	5,172 (24.0)	1,500 (36.7)	5,403 (22.8)	797 (60.6)
Externalb	1,696 (7.9)	333 (8.2)	1,773 (7.5)	115 (8.8)
Other	2,832 (13.2)	355 (8.7)	2,932 (12.4)	119 (9.1)

^aGTS = gentamycin and tobramycin sensitive, GNR = gram-negative rod, GTR = gentamycin and tobramycin resistant, AmS = amikacin sensitive, AmR = amikacin resistant.

^bAll data are number and percentage of isolates.

^c p < 0.001 for χ ² test.

Bivariate analyses

Selected bivariate associations are reported in Table 3. Male sex and black race were significantly associated with both GTR and AmR. The comorbidities most strongly associated with both GTR and AmR were cystic fibrosis and the Elixhauser score categories of weight loss and chronic pulmonary disease. In general, medical comorbidities were more strongly associated with GTR than with AmR, while measures of disease severity (longer length of stay, higher white blood cell count, presence of septic shock, and presence of invasive ventilation or devices) were more strongly associated with AmR than with GTR. Nonhematologic laboratory values were inconsistently associated with either resistance pattern. Consistent with prior literature, recent administration of aminoglycosides^14,20–22 and other antibiotics^15,23 was predictive of both GTR and AmR.

Table 3.

Selected Bivariate Associations of Patient Demographic and Clinical Factors With Resistance and Susceptibility Classification^a

Factor	GTS-GNR (n = 21,540)	GTR-GNR (n = 4,087)	p	AmS-GNR (n = 23,687)	AmR-GNR (n = 1,315)	p
Age, mean ± S.D.	64.2 ± 19.1	65.6 ± 18.7	<0.001	64.7 ± 19	62.4 ± 19	<0.001
Male sex	45.4	49.2	<0.001	44.6	56.7	<0.001
Race/ethnicity			<0.001			0.011
White	53.7	50.4		52.5	53.8
Asian	9.0	7.5		8.9	8.0
Black	11.1	13.7		11.5	13.0
Latino	20.0	20.8		20.8	17.6
Other	6.1	7.6		6.4	7.7
Body mass index	26.0 ± 6.7	25.5 ± 6.8	<0.001	26.1 ± 6.8	25.1 ± 6.6	<0.001
Admitted from healthcare facility,	14.2	28.5	<0.001	16.2	25.7	<0.001
Hospital (RRMC vs. SMH)	65.6	51.5	<0.001	63.2	53.6	<0.001
Log days to culture	0.44 (–1.51 to 2.11)	0.57 (–1.41 to 2.35)	<0.001	0.36 (–1.57 to 2.11)	1.32 (–0.58 to 2.69)	<0.001
Hospital-acquired infection	43.0	43.2	0.841	42.1	55.5	<0.001
In ICU at time of culture	18.4	18.7	0.656	17.8	25.8	<0.001
Any ICU stay during index hospitalization	36.2	41.6	<0.001	35.7	56.6	<0.001
Presence of indwelling urinary catheter	42.8	50.3	<0.001	43.8	65.8	<0.001
Ventilation during index hospitalization	28.6	38.7	<0.001	28.5	53.4	<0.001
Tracheostomy present on day of culture	10.5	17.8	<0.001	10.7	21.9	<0.001
Tracheostomy present on admission	4.5	10.8	<0.001	4.8	13	<0.001
Advanced ventilationc on day of culture	21.0	28.9	<0.001	20.8	41.8	<0.001
Elixhauser score	15 (6–26)	19 (9–29)	<0.001	16 (6–26)	21 (11–30)	<0.001
Comorbidities
Congestive heart failure	19.4	24.9	<0.001	20.3	25.6	<0.001
Arrhythmia	40.9	49.5	<0.001	42	52.4	<0.001
Valvular disease	23.4	27.2	<0.001	24	29.6	<0.001
Pulmonary vascular disease	15.6	19.1	<0.001	16.1	22.1	<0.001
Peripheral vascular disease	22.8	26.7	<0.001	23.2	28	<0.001
Paralysis	7.5	9.4	<0.001	7.9	8.7	0.292
Neurologic disease	27.6	37.9	<0.001	28.8	38.9	<0.001
Chronic pulmonary disease	23.6	31.6	<0.001	24.6	35.5	<0.001
Renal disease	32.2	39.2	<0.001	33.5	39.6	<0.001
Liver disease	23.8	25.4	0.021	24.5	27.9	0.005
Lymphoma	4.1	4.4	0.313	4.1	4.9	0.152
Metastatic cancer	10.6	8.2	<0.001	10.1	9.7	0.687
Nonmetastatic cancer	23.4	18.8	<0.001	22.8	20.6	0.067
Coagulopathy	25	30.7	<0.001	26	32.9	<0.001
Weight loss	17.2	25.7	<0.001	18	30.5	<0.001
Electrolyte disorder	58.7	66.8	<0.001	60.3	70	<0.001
Deficiency anemia	12.8	14.9	<0.001	13.3	13.8	0.635
Drug abuse	6.7	6.3	0.330	6.8	6.8	0.933
Depression	22.5	25.5	<0.001	23.4	25	0.179
Solid organ transplant	16.9	17.2	0.669	17.3	18.3	0.378
Bone marrow transplant	1.3	1.5	0.255	1.2	1.9	0.018
Renal failure	13.5	17.7	<0.001	14.5	18.3	<0.001
Cystic fibrosis	0.8	2.3	<0.001	0.7	5.5	<0.001
HIV infection	0.8	0.8	0.990	0.8	0.8	0.988
Alcohol use	23.3	17.3	<0.001	22	19.3	0.077
Tobacco use	5.9	5.1	0.123	5.7	4.5	0.154
Laboratory test values on day of culturec	91.2 ± 21.7	92.7 ± 22	0.002	91.3 ± 21.7	94.2 ± 21.7	<0.001
WBC count, cells x 109/L	12.4 (8.7–17.1)	12.7 (8.8–17.4)	0.061	12.3 (8.7–17)	13.4 (9.4–18.5)	<0.001
Hemoglobin, g/dL	10 (8.7–11.6)	9.7 (8.5–11.1)	<0.001	10 (8.7–11.6)	9.3 (8.3–10.6)	<0.001
Hematocrit, %	30.7 (26.8–35.3)	29.9 (26.3–34.1)	<0.001	30.0 (26.7–35.2)	29.0 (25.8–32.8)	<0.001
Platelet count, cells x 109/L	206 (134–289)	213 (129–302)	0.046	205 (133–288)	211 (119–314)	0.224
Sodium, meq/L	137 ± 6	138 ± 6	0.172	137 ± 6	138 ± 6	<0.001
Potassium, meq/L	4.1 ± 0.6	4.1 ± 0.6	<0.001	4.1 ± 0.6	4.1 ± 0.6	0.398
Chloride, meq/L	103 ± 6	102 ± 7	0.107	103 ± 7	102 ± 7	0.301
Bicarbonate, meq/L	24.4 ± 4.6	24.8 ± 5.1	<0.001	24.4 ± 4.6	25.5 ± 5.5	<0.001
Anion gap, meq/L	10.4 ± 4	10.3 ± 4.1	0.414	10.4 ± 4	10.1 ± 4.3	0.014
Creatinine, meq/L	1.4 ± 1.3	1.5 ± 1.4	<0.001	1.4 ± 1.3	1.4 ± 1.3	0.466
BUN, mg/dL	27.9 ± 22.4	32.2 ± 26.3	<0.001	28.4 ± 22.7	32.6 ± 28	<0.001
GFR, mL/min/1.73m2	71 (40–100)	68 (36–100)	<0.001	70 (39–100)	72 (37–100)	0.160
Glucose, mg/dL	135 ± 57	136 ± 61	0.480	135 ± 59	135 ± 54	0.686
Magnesium, mg/dL	1.7 ± 0.3	1.7 ± 0.4	0.006	1.7 ± 0.3	1.8 ± 0.4	<0.001
Calcium, mg/dL	8.6 ± 0.8	8.6 ± 0.9	0.565	8.6 ± 0.8	8.6 ± 1.1	0.248
Phosphorus, mg/dL	3.3 ± 1.2	3.4 ± 1.2	0.031	3.3 ± 1.1	3.4 ± 1.2	0.036
AST, IU/L	68 ± 352	65 ± 328	0.735	66 ± 335	77 ± 403	0.286
ALT, IU/L	51 ± 190	49 ± 167	0.503	50 ± 175	51 ± 156	0.846
ALK, IU/L	143 ± 161	153 ± 163	0.002	143 ± 156	164 ± 187	<0.001
aPTT, sec	22.1 ± 14.5	23.5 ± 15.5	<0.001	22.3 ± 14.9	24.8 ± 16.5	<0.001
INR	1.3 ± 0.6	1.4 ± 0.7	0.004	1.3 ± 0.6	1.4 ± 0.6	0.055
Lactate, mg/dL	20.1 ± 21.8	19.6 ± 20.2	0.338	20.3 ± 21.6	19.4 ± 21.8	0.275
Days since:
Last antibiotic	0 (0–9)	0 (0–4)	<0.001	0 (0–9)	0 (0–0)	<0.001
Last aminoglycoside	100 (100–100)	100 (100–100)	<0.001	100 (100–100)	100 (100–100)	<0.001
Last carbapenem	100 (100–100)	100 (47–100)	<0.001	100 (100–100)	100 (4–100)	<0.001
Last fluoroquinolone	100 (100–100)	100 (50–100)	<0.001	100 (100–100)	100 (21–100)	<0.001
Last penicillin	100 (2–100)	100 (3–100)	0.150	100 (3–100)	60 (1–100)	<0.001
Last anti-MRSA agent	100 (0–100)	24 (0–100)	<0.001	100 (0–100)	4 (0–100)	<0.001
Last colistin	100 (100–100)	100 (100–100)	<0.001	100 (100–100)	100 (100–100)	<0.001
Last aztreonam	100 (100–100)	100 (100–100)	<0.001	100 (100–100)	100 (100–100)	<0.001
Last β-lactam	0 (0–100)	0 (0–79.5)	0.089	0 (0–100)	0 (0–18)	<0.001
Last acid suppressant	0 (0–100)	0 (0–100)	<0.001	0 (0–100)	0 (0–2)	<0.001
Last probiotic	100 (100–100)	100 (100–100)	<0.001	100 (100–100)	100 (100–100)	<0.001
Last steroid	100 (60–100)	100 (22.5–100)	0.090	100 (61–100)	100 (4–100)	0.002
Last chemotherapy	100 (100–100)	100 (100–100)	0.233	100 (100–100)	100 (100–100)	0.015
Last immunosuppressant	100 (2–100)	100 (0.5–100)	0.312	100 (2–100)	100 (0–100)	0.021
Last blood product	100 (100–100)	100 (100–100)	0.001	100 (100–100)	100 (28–100)	<0.001

^aGTS = gentamycin and tobramycin sensitive, GNR = gram-negative rod, GTR = gentamycin and tobramycin resistant, AmS = amikacin sensitive, AmR = amikacin resistant, RRMC = Ronald Reagan UCLA Medical Center, SMH = Santa Monica UCLA Medical Center, ICU = intensive care unit, WBC = white blood cell, BUN = blood urea nitrogen, GFR = race-adjusted glomerular filtration rate, AST = aspartate aminotransferase, ALT = alanine aminotransferase, ALK = alkaline phosphatase, aPTT = activated partial thromboplastin time, INR = International Normalized Ratio.

^bData are percentage of patients, mean ± S.D., or median with interquartile range.

^cEither noninvasive mask ventilation or endotracheal intubation.

Multivariate analyses

There was substantial colinearity between many variables that were significant on bivariate analysis. In order to create a parsimonious model, these variables were tested against each other in groups, and the most representative predictors were used in further iterations of the model. These groups were as follows: medical comorbidities, demographics, vital signs, laboratory values, indwelling devices, and recently administered medications. Once the most representative predictors were chosen, they were added together, and the most parsimonious models were chosen. To facilitate model interpretability, the variables representing days since receipt of medications were dichotomized as receipt or nonreceipt within the prior 30 days; this did not significantly affect model fit. Hemoglobin was tested at multiple thresholds ranging from 7 to 13 g/dL; a cutoff of 11 g/dL was found to have the best model discrimination.

For the model predicting GTR-GNR, the predictors in the final model were presence of weight loss (as measured by the Elixhauser score category),²⁸ admission from another medical or long-term care facility, hemoglobin value of <11 g/dL, receipt of any carbapenem in the prior 30 days, and receipt of any fluoroquinolone in the prior 30 days; this model had a C statistic of 0.63 (Table 4).

Table 4.

Predictors of Aminoglycoside Resistance in Risk Prediction Models^a

Predictor	Coefficient	Standard Error	p
GTR-GNR Risk Prediction Model
Weight loss	0.41	0.07	<0.001
Facility prior to admission	0.84	0.06	<0.001
Hemoglobin <11 g/dL	0.23	0.06	<0.001
Carbapenems within 30 days	0.49	0.07	<0.001
Fluoroquinolones within 30 days	0.38	0.07	<0.001
AmR-GNR Risk Prediction Model
Cystic fibrosis	2.04	0.21	<0.001
Male gender	0.47	0.10	<0.001
Facility prior to admission	0.43	0.11	<0.001
Ventilation during hospitalization	0.70	0.10	<0.001
Carbapenems within 30 days	0.61	0.10	<0.001
Anti-MRSA agent within 30 days	0.53	0.10	<0.001

^aGTR = gentamycin and tobramycin resistant, GNR = gram-negative rod, AmR = amikacin resistant, MRSA = methicillin-resistant Staphylococcus aureus.

For the AmR-GNR prediction model, the predictors in the final model were somewhat overlapping: diagnosis of cystic fibrosis, male sex, admission from another medical or long-term care facility, ventilation at any point prior to culture during the index hospitalization, receipt of any carbapenem in the prior 30 days, and receipt of any anti-MRSA agent in the prior 30 days; this model had a C statistic of 0.74 (Table 4).

Treating each multivariate model as a score, with 1 point assigned for each of the items in the model, we created a user-friendly tool to predict the probability of GTR and AmR. Figures 1 and 2 show the positive predictive value at each score total for GTR and AmR, respectively, and demonstrate that a higher score is associated with a higher likelihood of resistance. Rates of GTR ranged from 10.2% at a score of 0 to 34% at a score of 4 or more. Rates of AmR ranged from 0.7% at a score of 0 to 17.3% at a score of 5; there were 0 cases with a score of 6. Due to the small number of cases with the highest score in each model, the 2 highest score values in each model were collapsed into a single category.

Figure 1.

Positive predictive value of the developed model for predicting gentamicin and tobramycin resistance of gram-negative rods at each score value.

Figure 2.

Positive predictive value of the developed model for predicting amikacin resistance of gram-negative rods at each score value.

Multinomial and ordinal logistic regression

The risk factors for the GTR-GNR and AmR-GNR models were combined to create an (unordered) multinomial logistic regression and an ordinal logistic regression model predicting the 3 categories of interest: aminoglycoside-susceptible, gentamicin- and tobramycin-resistant and amikacin-susceptible, and amikacin-resistant. By AIC, the multinomial model significantly outperformed the ordinal model. The Brant test of the proportional odds assumption indicated that the ordinal model did not appropriately describe the data. Taken together, these results indicated that the risk factors for GTR-GNR differ from those for AmR-GNR in kind and not merely degree.

Discussion

Ineffective empirical antibiotic choice is associated with poorer outcomes ^16–19; correct initial antibiotic choice is vital to reducing morbidity and mortality. Prior studies of risk factors for AmR among GNRs have largely focused on the family Enterobacteriaceae, which includes many commonly treated GNRs but excludes several clinically significant species, including Pseudomonas and Acinetobacter species. Additionally, many of these studies have been limited in scope (analyzing less than 200 patients), and none have resulted in creation of a predictive scoring system.^14,15,20,23 Our scores can be calculated by providers at the time of decision-making and, potentially, more accurately reflect a patient’s risk of aminoglycoside-resistant organisms relative to a hospitalwide or unit-specific antibiogram. All information used in the models was extracted directly from the medical record without any individual human examination of individual patient records, allowing for potential automated score calculation.

Results of our bivariate analysis were consistent with those of prior studies, confirming associations between AR and receipt of mechanical ventilation,^15,23 the presence of various indwelling devices,^15,23 longer hospital stay and more severe illness at the time of culture,²⁰ and recent exposure to various antibiotics.^{14,15,20,21,23} While it is improbable that exposure to all antibiotics mechanistically leads to a first AR occurrence, some of these exposures likely serve as a proxy for recent infection with MDR GNRs. We believe that at the study institutions, receipt of anti-MRSA agents is a proxy for recent concern for sepsis, as nearly all patients with suspected sepsis receive at least 1 dose of i.v. vancomycin; since there are other indications for receiving anti-MRSA medications besides empirical sepsis therapy, this is not a perfect proxy, but this phenomenon is likely responsible for the majority of the observed predictive power of our models. Results of our multivariate analysis suggest that factors associated with acute illness are less important than those associated with chronic illness and recent antibiotic exposure in determining the risk of AR-GNR occurrence.

The discrimination ability of the GTR-GNR model was more limited than that of the AmR-GNR model, likely due to the populations susceptible to GTS-GNR and GTR-GNR being relatively similar. GTR is substantially more common than AmR, and infections by resistant organisms are likely driven more by random chance than is acquisition of an AmR organism. The final model for AmR-GNR risk contains variables that are more relevant to chronic respiratory failure and prior severe infections, which are likely indicative of exposure to multiple antibiotics and repeated infections with organisms capable of developing first AmR. Additionally, results of the combined analysis indicate that the risk factors, while overlapping, are significantly different for the 2 resistance patterns. The model discrimination for the AmR-GNR model is substantially better than for the GTR-GNR model, indicating that the populations susceptible to AmR (versus AmS) GNR infections differ in more identifiable ways; thus, the model is better at ruling out AmR-GNR infections at lower scores, with our results indicating a less than 5% risk for AmR-GNR infection at scores under 3.

There were several limitations to our study. After eliminating duplicate cultures, approximately 60% of GTR and AmR-GNR cases could not be included in the final analysis due to a lack of complete data across the relevant domains. Additionally, since the data set only included information from inpatient hospitalizations, there was potentially lost predictive power due to some patients’ treatment at other facilities or outpatient encounters. We believe these limitations reflect the constraints on available data during real-world decision making and would be present with use of an automated version of the risk prediction tools we developed.

A significant potential limitation of our findings relates to external validity. Since all of the data came from 2 facilities of a single parent institution, it is possible that the developed models are not applicable to other medical systems. The work described involved several safeguards that mitigated this effect. First, the data set was derived from 2 hospitals that serve somewhat different populations. Ronald Reagan UCLS Medical Center has a high proportion of patients undergoing transplants of either solid organs or bone marrow, a robust neurosurgery patient population, and a large number of patients with chronic severe medical illness. Santa Monica UCLA Medical Center is a community hospital with a focus on geriatrics, orthopedic procedures, and solid oncology patients. Second, the data set represented tens of thousands of patients, a sample 1 to 2 orders of magnitude larger than those in similar studies examining risk factors for infection with resistant organisms. Third, as shown above, the bivariate predictor variables largely match up with predictors described in previous work. Finally, in order to prevent overfitting, the final models were restricted to a small number of predictor variables. Nevertheless, external validity is always a concern for single-institution studies, and further work will focus on validating these prediction scores in relation to more diverse data sets. Additionally, our investigation was the largest of its kind to date in terms of subject number and spanned a period of 6 years, allowing us to examine far more potential explanatory variables than prior investigations of risk factors for development of first GTR-GNR and AmR-GNR. By performing a cohort study of patients with positive cultures, we eliminated potential selection bias in choosing controls and strengthened the validity of observed associations.³⁰

While the discriminatory capacity of the GTR-GNR model is limited, the AmR-GNR model can effectively rule out AmR risk (i.e., there would be a risk of <5% at scores of <3), allowing reasonably confident treatment with amikacin as a first option without waiting for definitive amikacin susceptibility testing, which can take up to several days.

Our study demonstrated the potential power of data that is available to be harvested by automated methods from an existing electronic medical record. All information was extracted from standardized data fields via an automated process without any requirement for human interpretation or examination of individual records. The scoring system derived could be potentially integrated into an electronic health record and calculated automatically. In the current era of data-intensive medical care, we should harness all available information to better manage our patients. Further research will focus on validating the developed scoring method in other populations and analysis of cost–benefit thresholds for initiating specific antibiotic regimens in cases of uncertainty regarding pathogen identity.

Conclusion

A scoring system derived from the developed risk prediction models can be applied by providers to guide empirical antimicrobial therapy for treatment of GNR infections.

Disclosures

This work was supported by the National Institutes of Health (grant number U54GM114833-01) and the National Center for Advanced Translational Science (grant number UL1TR001881). The authors have declared no potential conflicts of interest.