Gut Dysbiosis Associated with Antibiotics and Disease Severity and Its Relation to Mortality in Critically Ill Patients

This study was conducted in the ICU of the Emergency Department of Osaka University Hospital, an academic urban tertiary referral hospital in Osaka, Japan, from February 2014 to January 2015. Patients who were intubated and required mechanical ventilation at admission or during their ICU stay were consecutively enrolled in this study. Exclusion criteria included postoperative patients with surgery of the rectum, patients with perianal infections, and patients expected to be on mechanical ventilation for less than 3 days. We included 71 patients in the study group, of whom 12 patients were enrolled in our previous study [14]. Informed consent for participation in this study was obtained from each patient or their next of kin. Nine healthy adult subjects were also included as a control group. They had never taken antibiotics and had not been admitted to hospitals within 3 months before participating in the study. This study was approved by the institutional review board of Osaka University Hospital (approval number: 12035).

In this observational prospective cohort study, fecal samples were collected serially from patients during hospitalization on days 1–2, 3–4, 5–7, 8–14, and thereafter when the researchers thought suitable. Fecal samples in the control group were collected on an arbitrary day without any symptoms. All samples were collected by inserting a sterile cotton-tipped swab 1–2 cm beyond the anus and rotating the swab for several seconds. Swabs were placed in sterile centrifuge tubes and immediately stored in a freezer at − 78 °C until use. DNA was extracted from these fecal samples using a DNeasy PowerSoil Kit (QIAGEN). To monitor changes in gut microbiota of the study population, 16S rRNA gene (V1–V2 region) deep sequencing was performed on a MiSeq system using a MiSeq v2 500 cycle kit (both, Illumina). The samples from each patient were sequenced in the same batch. The paired-end sequences obtained were merged, filtered, and denoised using DADA2. Taxonomic assignment was performed using the QIIME2 feature-classifier plugin with the Greengenes 13_8 database. The QIIME2 pipeline, version 2020.2, was used as the bioinformatics environment for the processing of all relevant raw sequencing data [15]. We collected and analyzed 238 samples of feces. Antibiotics class and duration were determined by the attending physician. Enteral nutrition with Glucerna®-Ex (Abbott Japan Co., Ltd.) was started via nasogastric tube within 48 h of admission if there were no contraindications. If infections occurred, patients were initially treated empirically for the underlying clinical syndrome using Gram staining and then according to the results of antibiotic susceptibility testing of the bacterial isolate causing the infection [16, 17]. Antibiotics were administered under the same policy during the entire study period. In trauma cases, antibiotics were administered prophylactically to prevent wound infections for about one week from admission.

Patient age, sex, Acute Physiology and Chronic Health Evaluation (APACHE II) score [18], and diagnosis were recorded on admission. The Sequential Organ Failure Assessment (SOFA) score [19] and the class of administered antibiotics were also recorded every day during hospitalization. Each patient’s outcome and date of death were also archived.

Patient baseline demographic and clinical variables are presented as median and interquartile range (IQR, 25th–75th percentile). The composition of the five major phyla (Bacteroidetes, Firmicutes, Proteobacteria, Actinobacteria and Fusobacteria) and each genus of gut microbiota are presented as percentages. The proportion of the five major phyla at the first sample was compared according to disease types (trauma, cardiac arrest, and sepsis) by Kruskal–Wallis test. To estimate the change in gut microbiota, we focused on two variables. One was the absolute proportional change in each phylum from the first sample to that on the sampling day to observe the overall changes from admission. The other was the relative proportional change of each phylum between the sampling day and the preceding sample to observe the daily changes (Additional file 1: Fig. S1). The composition rate of each phylum and the proportional changes at phylum level as well as the number of OTUs (operational taxonomic units: clusters of similar sequence variants of the 16S rRNA marker gene sequence, considered one of the milestones of biological diversity [α-diversity] [20]) of the healthy controls and the patient group on days 1–3, 4–7, and 8–14 were evaluated by Kruskal–Wallis test with the Dunn’s test used as a post hoc test. We also performed UniFrac distance analysis to compare the microbial flora structures [β-diversity] [21] between the healthy controls and the patient group on days 1–3, 4–7, and 8–14, and the distances were evaluated in the same way.

Carbapenems are thought to have one of the broadest spectrums of activity among all antibiotic classes; hence, we thought they would have the strongest influence on the gut microbiota. Samples were classified as belonging to the carbapenem group or the non-carbapenem group according to the antibiotic used on the day of sampling for a priori analyses. To account for dependencies in repeatedly measured observations within a subject, a nonlinear mixed-effect model was used with group, day, and interaction of group and day. Nonlinear associations between the difference in percentage in the phylum and day were assessed by including restricted cubic splines in the model. To establish the normality of the residuals of the mixed model and the nonlinear model, we performed a log transformation on the outcome. Because log conversion was not possible when the outcome was zero, we added a small enough value not to affect the results of the outcome before log conversion. Multivariable regression analyses were adjusted for age, sex, APACHE II score on admission, and SOFA scores. The absolute proportional changes of each phylum between the carbapenem group, penicillin group, and cephalosporin group were also estimated in the same way.

We also used a nonlinear mixed-effect model with initial SOFA score, SOFA score changes from the initial score on the sampling day, and interaction of initial SOFA score and SOFA score changes. Nonlinear associations between the difference in percentage in the phylum and SOFA score changes were assessed by including restricted cubic splines in the model. Multivariable regression analyses were adjusted for age, sex, and APACHE II score on admission. The absolute proportional changes of microbiota at each hospital day were tested using contrasts of the estimated regression coefficients in each nonlinear mixed effect model.

The changes in gut microbiota were also evaluated by the ratio of Bacteroidetes to Firmicutes (B/F ratio). Bacteroidetes and Firmicutes are the most predominant phyla in humans, and the B/F ratio is thought to be one of the indicators of change in the gut microbiota [22]. The maximum and minimum B/F ratios calculated during the study period were taken as representative values of the changes and of dysbiosis. We previously reported in a preliminary study that a B/F ratio > 10 or < 0.10 may be associated with poor prognosis [14]. The maximum and minimum B/F ratios within seven days from admission were analyzed by classification and regression tree (CART) analysis to explore the association with mortality. CART analysis is a binary recursive partitioning method using nonparametric approaches to identify high-risk and poor prognosis groups. The overall study group is split into two subgroups based on the most powerful predictor of the outcome [23]. Multivariate logistic regression analyses adjusted by sex and APACHE II score on admission were performed to evaluate the association of B/F ratio suggested by CART analysis. A receiver operating characteristic (ROC) curve was created to evaluate classification performance of B/F ratio suggested by CART analysis for the outcome of mortality, and the area under the curve (AUC) was calculated. Kaplan–Meier survival curves were constructed to display the probability of survival over the 56-day period in the patients with B/F ratios suggested by CART analysis and in those without them. The difference between them was evaluated by log-rank test.

All statistical inferences were made with a two-sided significance level of 5%. All statistical analyses were performed using SPSS (version 22, SPSS, Chicago, IL), R software version 3.3.2 using the “rms” package, and JMP Pro (version 11.2.0, SAS Institute, Cary, NC), and the survival curve graph was created using GraphPad Prism (version 6.04, GraphPad Software, La Jolla, CA).

Patient characteristics are shown in Table 1. During the study period, all patients required mechanical ventilation and received antibiotics intravenously. The antibiotic types are summarized in Table 1. Except for two patients, none of the patients were administered antibiotics within 3 months before study enrollment. None of the patients developed Clostridioides (Clostridium) difficile infection. Enteral nutrition was administered within 48 h from admission in 52 patients (73.2%), and 15 patients (21.1%) did not receive enteral feeding during the study period. The median duration of ICU stay was 17 (IQR: 9–37) days. In this study, 24 patients (34%) died from causes of death that included multiple organ failure in 10 patients, cardiac failure in 8 patients, and brain death in 6 patients.

All sampled data are summarized in Additional file 2: Table S1. The taxonomic composition of the gut microbiota at phylum level of the first sample and the last sample in each patient and the healthy controls is shown in Fig. 1. Bacteroidetes and Firmicutes were the predominant phyla in most patients, and their proportions varied from patient to patient at the first sample. The composition of the gut microbiota dramatically changed during the ICU stay. The occupancy rates of the top 10 taxonomic compositions at phylum, class, order, family, and genus level of all patient samples are shown in Additional file 1: Fig. S2. The five major phyla in the patients were Bacteroidetes (44.1%), Firmicutes (34.1%), Proteobacteria (12.7%), Actinobacteria (4.2%), and Fusobacteria (3.1%). The proportion of each phylum at the first sample between patients with trauma, cardiac arrest, and sepsis is shown in Additional file 1: Table S2. There were no significant differences between diseases types except for Proteobacteria (p = 0.032).

The composition rate and absolute proportional changes of each phylum as well as changes in the number of OTUs in the healthy controls and the patient group on days 1–3, 4–7, and 8–14 from admission are shown in Fig. 2. The composition rate of each phylum was not significantly different between each sampling day. However, the absolute proportional differences of Bacteroidetes and Firmicutes significantly increased over time (p < 0.001; Fig. 2a, b), and the number of OTUs significantly decreased over time (p < 0.001; Fig. 2c). To compare the microbial community between the healthy controls and the patient group on each sampling day, we also performed UniFrac distance analysis of our 16S rRNA gene sequencing data. The distance between the healthy controls and the patient group on each sampling day significantly increased until day 7 (p < 0.05; Fig. 2d), which meant that the microbial flora structures gradually changed until day 7. The taxonomic composition of the gut microbiota at genus level during the ICU stay and that of the healthy controls is shown in Additional file 1: Fig. S3. There were significant changes in the proportions of Blautia, Clostridium, Coprococcus, Enterococcus, Faecalibacterium, Lacnospiraceae, Roseburia, Ruminococcus, and Streptococcus between the healthy controls and each sampling day.

These results suggested that each ICU patient had their own unique gut microbiota at admission, and the dysbiosis was gradually proceeding with the loss of diversity of gut microbiota during the acute phase of the ICU stay.

To assess the association between the use of broad-spectrum antibiotics and the gut microbiota of the ICU patients, the absolute proportional change from the first sample and the relative proportional change from the preceding sample were estimated between the carbapenem and non-carbapenem groups by use of a mixed-effect model.

The absolute proportional changes of Bacteroidetes and Firmicutes increased and stabilized approximately on the 7th hospital day in both groups (Fig. 3). Statistically significant differences were found between the carbapenem and non-carbapenem groups in Bacteroidetes from the 11th to 15th hospital day and in Firmicutes from the 7th to 14th hospital day (p < 0.05). The relative proportional changes of Bacteroidetes and Firmicutes converged almost to zero approximately on the 7th hospital day in both groups (Fig. 4).

Among the other phyla, Actinobacteria showed a significant difference in the absolute proportional change between the carbapenem and non-carbapenem groups (p = 0.030; Fig. 3). Significant differences in the absolute proportional changes in Proteobacteria from the 1st to 4th hospital day, in Actinobacteria from the 13th to 17th hospital day, and in Fusobacteria from the 14th to 17th hospital day (p < 0.05) were also found by testing for contrasts in interaction (Fig. 3). The relative proportional change of Proteobacteria converged almost to zero approximately on the 7th hospital day in both groups (Fig. 4).

To identify differences between other antibiotics, the absolute proportional changes of each phylum between the carbapenem group, penicillin group, and cephalosporin group were also estimated (Additional file 1: Fig. S4). Each phylum reached inflection points approximately on the 7th hospital day in each antibiotics group, but the differences did not reach statistical significance between the groups.

These results suggested that the major changes of the gut microbiota, especially of Bacteroidetes and Firmicutes, may be completed within the first week after admission to ICU, and broad-spectrum antibiotics may be associated with the change of the composition of gut microbiota from the early phase after ICU admission. Actinobacteria may be more sensitive to carbapenems than other phyla.

To estimate the association between disease severity and the gut microbiota of the ICU patients, the absolute proportional changes from the first sample were estimated by a linear mixed-effect model with interaction of initial SOFA score and changes in SOFA score. There was a statistically significant difference in the absolute proportional change in Actinobacteria with interaction of the initial SOFA score and changes in SOFA score during the patients’ ICU stay (p < 0.001; Fig. 5). This result suggested that disease severity may be associated with the change in the proportion of Actinobacteria during the acute phase of ICU stay.

We explored whether gut dysbiosis during the acute-phase ICU stay could be associated with mortality by using CART and multivariate logistic regression analyses. The cutoff values of the maximum and minimum B/F ratios within seven days of admission were determined to be 8 and 1/8, respectively. The diagnostic characteristics of this tree have a sensitivity of 54.2% and specificity of 74.5% (Fig. 6a). ROC curve analysis using these B/F ratios for the outcome of ICU mortality was performed, and the AUC was 0.853 (Fig. 6b). Differences in the rates of ICU mortality were calculated by Kaplan–Meier analysis using these B/F ratios, and the hazard ratio was 2.41 (95% confidence interval [CI]: 1.04–5.51, p = 0.039) (Fig. 6c). Multivariate logistic regression analysis showed an association between an imbalance in B/F ratio and higher mortality when the B/F ratio was > 8 or < 1/8 (odds ratio: 5.54, 95% CI: 1.39–22.18, p = 0.015; Table 2). We also performed multivariate logistic regression analysis adjusted for APACHE II score on admission, sex, and B/F ratio (> 8 or < 1/8) of the fecal samples on the first three days from admission or at seven to nine days from admission to estimate ICU mortality (Additional file 1: Table S3). The B/F ratio on the first three days from admission was not associated with ICU mortality, whereas the B/F ratio one week after admission was associated with ICU mortality.

These results suggested that the B/F ratio within seven days from admission might be useful as an indicator of poor prognosis and that an extreme change in the composition of the gut microbiota may be associated with mortality in the ICU. Although it may be difficult to predict prognosis based on the value of the B/F ratio on admission, changes in the B/F ratio during the ICU stay may be more useful for the prediction of prognosis.