Effect of delays in concordant antibiotic treatment on mortality in patients with hospital-acquired Acinetobacter spp. bacteremia: a 13-year retrospective cohort

Background Therapeutic options for multidrug-resistant Acinetobacter spp. are limited, and resistance to last resort antibiotics in hospitals is increasing globally. Quantifying the impact of delays in concordant antibiotic treatment on patient mortality is important for designing hospital antibiotic policies. Methods We included patients with Acinetobacter spp. hospital-acquired bacteremia (HAB) in a hospital in Thailand over a 13-year period. For each day of stay following the first positive blood culture we considered antibiotic treatment to be concordant if the isolated organism was susceptible to at least one antibiotic given. We used marginal structural models with inverse-probability weightings to determine the association between delays in concordant treatment and 30-day mortality. Results Between January 2003 and December 2015, 1,203 patients had HAB with Acinetobacter spp., of which 682 patients (56.7%) had one or more days of delay in concordant treatment. These delays were associated with an absolute increase in 30-day mortality of 6.6% (95% CI 0.2%-13.0%), from 33.8% to 40.4%. Crude 30-day mortality was substantially lower in patients with three or more days of delays in concordant treatment compared to those with one to two days of delays. Accounting for confounders and immortal time bias resolved this paradox, and showed similar 30-day mortality for one, two and three or more days of delays. Conclusions Delays in concordant antibiotic treatment were associated with a 6.6% absolute increase in mortality among patients with hospital-acquired Acinetobacter spp. bacteremia. If this association is causal, switching fifteen patients from discordant to concordant initial treatment would be expected to prevent one death.

CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.   CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.   CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 3, 2020.

84
Randomized controlled trials are the "gold standard" design to study the causal relationship between a 85 treatment and an outcome. However, it would be unethical to intentionally delay the provision of 86 concordant antibiotic treatment to a patient. Therefore, an observational cohort study is likely to be the 87 best way to determine the impact of delays in concordant antibiotic treatment on patient outcomes. There  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 3, 2020. . https://doi.org/10.1101/2020.04.27.20081513 doi: medRxiv preprint 5 an infection, and these changes will often be related to the time-varying clinical severity of the patient, 96 which could be a confounder for the current treatment and a mediator for the future treatment (Box 1).

97
Finally, patients may die or be discharged from the hospital before the empirical treatment is completed, 98 and excluding these patients from the final analysis could introduce immortal time bias. 11-13 99 100 In this study, our objective is to study the relationship between days of delay in concordant antibiotic 101 treatment and patient mortality by performing an analysis on observational data accounting for these 102 distorting factors, aiming, as far as possible, to emulate a randomized trial.

103
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 3, 2020.

106
We identified, from a 13-year retrospective cohort, patients with hospital-acquired bacteremia (HAB) 107 related to Acinetobacter spp. in Sunpasitthiprasong Hospital, Thailand. This is a provincial hospital with 108 1,500 beds. The hospital has a microbiology laboratory that performs culture and isolate identification

127
We considered any antibiotics prescribed within three days of the first blood sample collected and with 128 Acinetobacter spp. isolated to represent empirical treatment. This is because microbiological 129 identification and antibiotic susceptibility testing usually require three days. An antibiotic regimen was 130 defined as concordant if susceptibility testing indicated that the isolated organism was susceptible to at 131 least one of the antibiotics given. Otherwise the regimen was defined as discordant. Concordance of the 132 antibiotic treatment was determined for each eligible patient on the day of blood sample collection (t=0),

133
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020. . https://doi.org/10.1101/2020.04.27.20081513 doi: medRxiv preprint 8 The effects of delays in concordant empirical antibiotic treatment on 30-day mortality were estimated 163 using marginal structural models. 20 We performed two analyses. The first analysis was to evaluate the 164 impact of one or more days of delays in concordant antibiotic treatment. A propensity score for each 165 patient was calculated to represent the probability of being prescribed with concordant antibiotic on the 166 day of blood collection using a logistic regression model adjusting for the pre-specified confounders and 167 using fractional polynomials to account for non-linear relationships. The propensity scores were then 168 used to calculate stabilized inverse probability weights (IPW), which were applied to a marginal 169 structural model. The second analysis was to evaluate the effect of one-day, two-day, and three or more

192
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.     CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.    (Table 2). While the crude 240 analysis suggested that patients with three or more days of delays in concordant antibiotic treatment had 241 the lowest mortality, the adjusted analysis found that the expected mortality was the lowest if the patients 242 had no delays, though found no evidence of increasing mortality with increasing delays. However, 243 uncertainty was large. Absolute differences in mortality between no delays in concordant antibiotic 244 treatment and a one-day delay, a two-day delay, and three or more days of delays were 3.0% (95% CI -245 12.0%-18.0%); 11.3% (95% CI -3.0%-25.6%); and 1.1% (95% CI -7.8%-10.0%), respectively. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.

261
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.

262
After adjusting for measured confounders, we found that delays in concordant antibiotic treatment of one 263 or more days were associated with an absolute increase of 6.6% in 30-day mortality from 33.8% to 264 40.4%. If this increase is caused by the delay in concordant treatment, it would indicate that we would 265 need to switch empirical antibiotic treatment from discordant to concordant in fifteen patients in order to 266 prevent one 30-day death. There was no evidence of a dose response relationship between number of 267 days of delays in concordant antibiotics and 30-day mortality.

269
The immortal time bias observed in this cohort is a common phenomenon, and needs to be considered in

303
In most previous studies antibiotic use has been considered as a binary variable and switching of 304 antibiotic regimens due to changes in clinical symptoms has been neglected. 31 In hospitals in resource-305 limited settings, antibiotics are sometimes prescribed even before a clinical specimen is taken for culture, 306 and switching regimen in response to severity of infection is common. 31 This change in regimen 307 determined by clinical symptoms, if not adjusted for using appropriate methods for time-varying 308 confounders, may also lead to biases. 20 Marginal structural models have been used to adjust for time-

309
varying confounders in a previous study of the association between appropriate antibiotic treatment for 310 bacteremia on the day the blood culture was taken and mortality and discharge. 32 The previous study 311 found no evidence for a protective effect of appropriate empirical antibiotic treatment on mortality and 312 discharge, but confidence intervals were wide. 32 Differences in bacterial species considered, patient 313 characteristics and clinical setting make direct comparison with the current study inappropriate.

315
Our study has limitations. Firstly, data on severity of infections were not routinely collected, and this is 316 typically the case in LMIC settings. We used admission to ICU and prescription of vasopressor as proxy 317 variables for the severity of infection. These proxy variables will only imperfectly represent the severity 318 of infection and residual confounding is to be expected; however, both are specific in representing severe 319 clinical conditions. Secondly, despite the relatively large sample size, the power to detect a dose-response 320 relationship in the four regimens under evaluation might be low. This is reflected in the wide confidence 321 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.

325
In conclusion, we observed a 6.6% absolute increase in mortality among patients with hospital-acquired 326 Acinetobacter spp. bacteremia when concordant antibiotic treatment was delayed for one or more days.

327
Accounting for confounding and immortal time biases is necessary when attempting to estimate causal 328 effects of delayed concordant treatment and, in this case, helped resolve paradoxical results in crude data. 329 330 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted May 3, 2020.

359
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 3, 2020.

387
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 3, 2020.  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 3, 2020.

449
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 3, 2020. . https://doi.org/10.1101/2020.04.27.20081513 doi: medRxiv preprint