Clinical and Culture-Based Predictors of Gut Microbiome Alpha Diversity at the Time of ICU Admission

Abstract

OBJECTIVES:

Microbiome-based therapies to improve gut colonization resistance are being developed for the ICU, but their efficacy may depend on the baseline gut microbiome at ICU admission. We sought to identify clinical predictors of alpha diversity at ICU admission.

DESIGN:

Retrospective reanalysis of a randomized clinical trial (NCT03865706).

SETTING:

Single-center ICU.

PATIENTS:

ICU patients with sepsis as the primary diagnosis who were receiving broad-spectrum antibiotics and could be enrolled within 24 hours of ICU admission.

INTERVENTIONS:

None.

MEASUREMENTS AND MAIN RESULTS:

Demographic and clinical characteristics for medical ICU patients with sepsis were recorded during a previously published randomized clinical trial. Deep rectal swabs were collected within 24 hours of ICU admission and sequenced to describe alpha diversity (Shannon index) and cultured for vancomycin-resistant Enterococcus (VRE). Patients were organized into tertiles of Shannon diversity (low, middle, high) with a primary outcome of a lower Shannon tertile at ICU admission. Overall, 90 patients were enrolled. The three variables of location before ICU admission (adjusted odds ratio [aOR], 3.54; 95% CI, 1.21–10.3 for ICU transfer vs. emergency department [ED]; aOR, 6.09; 95% CI, 1.89–19.6 for hospital ward vs. ED), prior culture-proven infection within 1 year (aOR, 2.46; 95% CI, 1.02–5.96), and VRE status on ICU admission (aOR, 4.18; 95% CI, 1.44–12.1 for positive vs. negative swab) were sufficient to describe a quasi-linear trend in alpha diversity at ICU admission.

CONCLUSIONS:

Baseline gut microbiome Shannon alpha diversity at ICU admission could be described with three readily ascertained clinical variables. Our model shows promise as a preliminary framework of factors that collectively best predict baseline alpha diversity at ICU admission and warrants validation in larger independent cohorts.

KEY POINTS.

Question: Which clinical risk factors are associated with a low baseline gut microbiome diversity on ICU admission?

Findings: The three variables of location before ICU admission, prior culture-proven infection within 1 year, and vancomycin-resistant Enterococcus status on ICU admission were sufficient to describe baseline gut microbiome Shannon alpha diversity at ICU admission.

Meaning: Our simple model offers a hypothesis-generating framework to rapidly determine a patient’s gut microbiome alpha diversity on ICU admission, pending validation in larger independent cohorts.

The gut microbiota of patients hospitalized in the ICU demonstrates frequent gut pathogen colonization and is associated with higher downstream rates of infection (1–3). Baseline microbiome composition at ICU admission is likely to be an important determinant of gut microbiome throughout hospitalization and therefore may predict clinical prognosis, with low baseline fecal microbial diversity representing loss of colonization resistance (4–7). Therapies are actively being developed for patients in the ICU that seek to improve gut colonization resistance and infection outcomes. However, the efficacy of microbiome-based therapies in ICU patients may depend on the baseline diversity of the microbiome, as patients with an already robust microbiome would presumably be less likely to benefit from microbiome restitution (8).

Prior studies have examined clinical risk factors for gut colonization with specific pathogens in the ICU. Relatively few studies have taken a more holistic approach or incorporated stool culture and sequencing results to analyze risk factors for low fecal microbial diversity. Regression models have been developed to predict multidrug-resistant organism (MDRO) colonization status using clinical variables, particularly for vancomycin-resistant Enterococcus (VRE) colonization at admission. Risk factors for VRE carriage include recent antibiotic treatment, previous hospitalizations, increased length of stay, and comorbid conditions (9–12). Gut colonization with MDROs has been associated with lower alpha diversity at the time of ICU admission (10, 13, 14).

In addition to MDRO cultures, clinical factors have been associated with reduced microbiome alpha diversity. For example, antibiotics and medication use have been shown to influence gut microbiome composition (2, 15–19). A history of ICU hospitalizations, sepsis, and chronic medical comorbidities are also associated with altered microbiome composition at ICU admission (17, 20–22). Furthermore, reduced gut microbiome alpha diversity is a predictor of poor ICU outcomes in prior studies including prolonged hospitalization, increased complications, infections, and organ failure, and higher mortality (23, 24). As patients with reduced diversity are at a higher risk for these adverse clinical outcomes in the ICU, they represent a clinically important target population for microbiome-based interventions.

Despite this prior work, there remains uncertainty regarding the clinical factors that collectively best predict low microbiome alpha diversity at the time of ICU admission, the strength of these predictors, and the extent to which standard MDRO cultures may add to clinical predictors for low baseline diversity. While culture data may not be immediately available at the time of admission, cultures could be done promptly, so stool culture could plausibly be used as part of trial entry criteria for future microbiome therapeutics alongside clinical variables. In sum, there is a need to understand the clinical predictors of low fecal diversity in this setting, which would allow more efficient selection of the patients who would be most likely to benefit from microbiome-based interventions and be included in future clinical trials. We sought to determine which clinical and culture-based factors best predicted low fecal microbial diversity at the time of admission to the medical ICU for patients with sepsis who may be candidates of future trials.

METHODS

Population

The study cohort was drawn from a published single-center, double-blind, placebo-controlled randomized clinical trial (RCT) of prebiotic inulin in preventing gut pathogen colonization and infection in the ICU (NCT03865706) (4). The trial was registered through ClinicalTrials.gov on March 4, 2019, and approved by the Columbia University Institutional Review Board (No. AAAS2576) on April 5, 2019, under the study title “Inulin in the ICU.” Informed consent was obtained from all study participants and procedures were followed in accordance with the ethical standards of the responsible committee on human experimentation and the Helsinki Declaration of 1975. The primary outcome of the trial was null, and the trial data were reutilized for this study. The trial sought to enroll a standard medical ICU population generalizable to all medical ICUs. Participants were patients admitted to the ICU with sepsis as the primary diagnosis who were receiving broad-spectrum antibiotics and could be enrolled within 24 hours of ICU admission. Exclusion criteria included testing positive for COVID-19, inability to receive enteral fluids, or limited treatment goals such as “do not resuscitate” or “no escalation of care” orders.

Samples

Two protocolized deep rectal swabs were collected by a study nurse from all participants at the time of ICU admission. Rectal swabs were performed with the patient in the left lateral decubitus position, with fecal soilage used to ensure adequate sampling. One of the rectal swabs was flash frozen at –80ºC for 16S ribosomal RNA (rRNA) gene sequencing, and the other swab was added to soy broth and cultured for VRE and other MDRO bacteria.

16S rRNA Gene Sequencing

Sequencing was performed from extracted DNA from rectal swabs as previously described (4). Sequencing data are publicly available through the National Center for Biotechnology Information Sequencing Read Archive (accession number PRJNA1128691). Shannon alpha diversity scores were calculated from sequencing data using the phyloseq package on R and organized into tertiles of low, medium, and high diversity gut microbiomes (25). The Shannon index H is calculated as $H=-\underset{i=1}{\sum ^S}{p}_i\mathrm{l}\mathrm{n}\left(p_i\right)$, where S is the total number of species and p_i is the proportion for the i^th species. Although various metrics of diversity exist, the Shannon index was chosen as it is most widely used and accounts for both species richness and evenness in a single parameter (26, 27).

Bacterial Culture

Duplicate rectal swabs collected on ICU admission were inoculated directly onto chromogenic agar selective for VRE, extended-spectrum beta-lactamase-producing Gram-negative (ESBL-GN) bacteria, and carbapenem-resistant Enterobacteriaceae (CRE) and incubated for up to 3 days at 37°C. Species identification was obtained using full-length 16S rRNA gene sequencing on picked colonies from blood agar plates (VRE) and MacConkey plates (ESBL-GN).

Clinical Data Collection

Demographics, Charlson comorbidity scores, Sequential Organ Failure Assessment scores, and other ICU admission characteristics were recorded prospectively during the clinical trial. Additional variables were collected retrospectively via chart review in the Epic-based electronic medical record (EMR). Prior hospitalization records within and outside of our institution within 1 year of medical ICU admission were reviewed, with hospitalization defined as any hospital admission longer than one night. Data on the number, duration, reason, and location of hospitalizations were analyzed. Prior infections were identified by searching historical medical records for any positive microbial cultures with corresponding clinical documentation of symptoms and signs consistent with infection within 1 year of ICU admission and did not include colonization without symptoms. Medication records within 90 days of ICU admission were obtained through a review of all medication lists from the EMR. The number, route of administration, classes, and duration of use of antibiotics were recorded, along with use of other medications such as nonsteroidal anti-inflammatory drugs, proton pump inhibitors, beta blockers, metformin, angiotensin-converting enzyme inhibitors, laxatives, antidepressants, and corticosteroids. The duration of antibiotic use was further divided into unique days of antibiotic use, defined as the number of calendar days of antibiotic exposure, and total days of antibiotic use, which refers to the aggregate sum of durations for all antibiotics used. For patients who did not receive antibiotics within 90 days of ICU admission, the date of their last antibiotic use was recorded so that we could calculate the time from their last antibiotic use to ICU admission.

Statistical Analysis

The primary outcome was the Shannon alpha diversity index at ICU admission, which was organized into tertiles consisting of 30 subjects each. We elected not to model the Shannon index as a continuous measure due to the difficulty in interpreting continuous outcomes and to avoid assumptions regarding the data distribution and linearity. The clinical and culture-based variables were then compared across tertiles to determine the relationship between these factors and baseline gut alpha diversity. Demographic data were summarized using descriptive statistics. The chi-square tests were used to compare categorical measures, and the Fisher exact test was used for categorical measures with low expected counts (< 5). Kruskal-Wallis rank-sum tests were used to compare continuous measures, and the Dwass-Steel-Critchlow-Fligner test was used for post hoc pairwise comparisons with continuous measures. Data were analyzed in RStudio, Version 2025.09.0, Jamovi, Version 2.6.26.0, and STATA, Version 15.

Model Construction

An ordinal logistic regression model was constructed to test the independent effects of predictor variables. The model outcome was a lower tertile of Shannon index (i.e., moving from tertile 3 to tertile 2, or moving from tertile 2 to tertile 1). Initially, we examined all predictor variables with an association with Shannon tertile (p < 0.10). Variables were then organized into categories to eliminate collinearity between variables. For example, unique antibiotic classes and unique days of antibiotics both fall within the broader category of antibiotic use. Within categories, we used the Akaike information criterion (AIC) to select one variable per category for the final, reduced model to avoid over-parameterizing the model. This was done for the categories of prior infection, prior healthcare exposure, and prior antibiotics. Last, we selected the top three predictor variables most closely associated with Shannon tertile at ICU admission that produced the best model fit and visualized their relationship with Shannon index when it was reclassified as a continuous measure.

RESULTS

Participant Demographics and Lower Alpha Diversity

Ninety medical ICU patients with sepsis were organized into tertiles of Shannon alpha diversity at ICU admission, with 30 patients per tertile. The median Shannon diversity index for the low, medium, and high tertiles of alpha diversity were 1.17 (interquartile range [IQR], 0.42–1.68), 2.79 (IQR, 2.67–3.00), and 3.52 (IQR, 3.35–3.68), respectively. Demographic characteristics were similar across tertiles (Table 1).

TABLE 1.

Relationship Between Patient Characteristics and Alpha Diversity at ICU Admission

Characteristics	Alpha Diversity			p a
	Tertile 1, Low (n = 30)	Tertile 2, Medium (n = 30)	Tertile 3, High (n = 30)
Shannon index, median (IQR)	1.17 (0.42–1.68)	2.79 (2.67–3.00)	3.52 (3.35–3.68)	< 0.01
Demographics at ICU admission
Age, yr, median (IQR)	63.5 (55.3–71.8)	66.5 (57.3–74.3)	58.5 (40.3–69.3)	0.19
Sex				0.25
Male	13 (43)	19 (63)	18 (60)
Female	17 (57)	11 (37)	12 (40)
Race/ethnicity				0.63
White	13 (43)	13 (43)	9 (30)
Black	6 (20)	5 (17)	8 (27)
Hispanic	7 (23)	7 (23)	9 (30)
Asian	0	2 (7)	0
Other	4 (13)	3 (10)	4 (13)
Clinical characteristics at ICU admission
Location before ICU admission				0.01
Emergency department	13 (43)	19 (63)	26 (87)
Hospital ward	12 (40)	5 (17)	3 (10)
Another ICU	5 (17)	6 (20)	1 (3)
Charlson comorbidity score, median (IQR)	4 (3–5)	4 (2.25–6)	2 (0.25–5.75)	0.04
Sequential Organ Failure Assessment score, median (IQR)	9 (7–11)	8 (7–10)	8 (6–11)	0.28
Mechanical ventilation	23 (77)	21 (70)	28 (93)	0.07
Use of vasopressors	24 (80)	20 (67)	20 (67)	0.47
Culture-proven infections, within the year before ICU admission
Any culture-proven infection	24 (80)	21 (70)	15 (50)	0.04
Any multidrug-resistant organism infection	14 (47)	13 (43)	6 (20)	0.07
Vancomycin-resistant Enterococcus infection	5 (17)	4 (13)	0	0.07
Extended-spectrum beta-lactamase-producing Gram-negative infection	12 (40)	10 (33)	4 (13)	0.06
Methicillin-resistant Staphylococcus aureus infection	3 (10)	3 (10)	2 (7)	1.00
Healthcare exposure, within the year before ICU admission
Any hospitalization	22 (73)	24 (80)	19 (63)	0.35
Number of hospitalizations, median (IQR)	1 (0.25–3.00)	2 (1.25–4.00)	1 (0–2.00)	0.06
Total days in hospital, median (IQR)	13 (3.5–60)	29 (12–63)	3.5 (0–15)	< 0.01
Any ICU hospitalization	9 (30)	8 (27)	5 (17)	0.46

IQR = interquartile range.

^aχ² test was used for categorical measures, Fisher exact test was used for categorical measures with low (< 5) expected counts, and the Kruskal-Wallis rank-sum test was used for continuous measures.

Data are presented as n (%) of patients, except where indicated.

Clinical Characteristics and Lower Alpha Diversity

Next, we examined clinical characteristics as predictors of alpha diversity. The location of patients before the ICU admission varied significantly between tertiles, with the low diversity patients more likely to be transferred from the hospital ward or another ICU rather than coming through the emergency department (ED; Table 1). Low diversity patients also had significantly higher median Charlson comorbidity scores compared with high diversity patients. Specific classes of chronic comorbidities (e.g., cardiovascular, lung, liver) were also investigated but showed no consistent difference across tertiles. Within the year before ICU admission, patients with lower alpha diversity had significantly more culture-proven infections (Table 1). These patients were also more likely to have had MDRO infections, although these differences were not statistically significant. The total number of days in the hospital before ICU admission differed significantly between the tertiles, but the difference was mostly driven by the high median in tertile 2. We isolated the patients who had been hospitalized in each tertile and reanalyzed the hospitalization characteristics among this subset, with similar results (Supplemental Table 1, https://links.lww.com/CCX/B666).

Antibiotic Use and Lower Alpha Diversity

Antibiotic use is an established risk factor for low gut alpha diversity, so we interrogated the relationship between different measures of antibiotic use and alpha diversity scores on ICU admission. We found that patients with lower alpha diversity scores were more likely to have used antibiotics, used increased number and classes of antibiotics, and had increased total and unique days of antibiotic use in the 3 months before ICU admission, compared with patients with high alpha diversity (Fig. 1A and Table 2). However, there was not always a clear dose-response effect across tertiles. Results were similar, without clear correlations, when we examined alpha diversity as a continuous measure (Supplemental Figs. 1 and 2, https://links.lww.com/CCX/B666). There was a modest relationship between continuous Shannon diversity and time since last antibiotic use (Fig. 1B).

Figure 1.

Shannon alpha diversity and antibiotic usage. A, Shannon alpha diversity and antibiotic usage within 3 mo before ICU admission. This panel does not classify patients as antibiotic exposed if they used antibiotics more than 3 mo before ICU admission. p value was calculated using Kruskal-Wallis rank-sum test for continuous measures. B, Shannon alpha diversity and duration since last antibiotic use. Patients who last used antibiotics more than 3 mo before ICU admission were included in this figure, but those who last used antibiotics on the day of ICU admission were excluded. Trendline was plotted with the second order polynomial line of best fit generated through R.

TABLE 2.

Relationship Between Antibiotic Use 3 Months Before ICU Admission and Alpha Diversity at ICU Admission

Medications	Alpha Diversity			p a
	Tertile 1, Low (n = 30)	Tertile 2, Medium (n = 30)	Tertile 3, High (n = 30)
Antibiotics
Number of antibiotics, median (IQR)	4.00 (2.25–5.75)	4.00 (2.25–5.00)	2.00 (0–4.75)	0.05
Number of unique antibiotic classes, median (IQR)	4.00 (2.25–5.00)	3.00 (2.25–4.75)	2.00 (0–4.00)	0.05
Total days of antibiotic use, d, median (IQR)	11 (3–46)	16.5 (4–43)	2 (0–19)	0.13
Unique days of antibiotic use, d, median (IQR)	6 (1–34)	11 (3–23)	1 (0–12)	0.04
Specific antibiotic classes used
Aminoglycosides	5 (17)	2 (7)	1 (3)	0.26
Carbapenems	9 (30)	8 (27)	3 (10)	0.14
Cephalosporins	16 (53)	16 (53)	12 (40)	0.49
Fluoroquinolones	9 (30)	4 (13)	2 (7)	0.04
Glycopeptides	21 (70)	18 (60)	13 (43)	0.11
Lincosamides	0	0	4 (13)	0.03
Macrolides	8 (27)	7 (23)	10 (33)	0.68
Monobactams	3 (10)	0	2 (7)	0.36
Nitrofurantoin	0	2 (7)	1 (3)	0.77
Nitroimidazole	4 (13)	5 (17)	1 (3)	0.33
Oxazolidinone	4 (13)	1 (3)	0	0.12
Penicillins	19 (63)	23 (77)	12 (40)	0.01
Phosphonics	0	2 (7)	0	0.33
Polymyxins	0	1 (3)	1 (3)	1.00
Rifamycin	3 (10)	0	1 (3)	0.32
Sulfonamides	7 (23)	5 (17)	4 (13)	0.59
Tetracycline	7 (23)	2 (7)	2 (7)	0.10
Other	1 (3)	0	1 (3)	1.00

IQR = interquartile range.

^aχ² test was used for categorical measures, Fisher exact test was used for categorical measures with low (< 5) expected counts, and the Kruskal-Wallis rank-sum test was used for continuous measures.

Data are presented as n (%) of patients, except where indicated.

We further explored the relationship between the number of antibiotic classes used in the 3 months before ICU admission and alpha diversity (Supplemental Fig. 3, https://links.lww.com/CCX/B666). Patients who used three or more classes of antibiotics had significantly lower alpha diversity than patients who used two or fewer classes of antibiotics. Examining specific antibiotic classes, we found that lower diversity was weakly associated with use of fluoroquinolones and penicillins (Table 2 and Supplemental Fig. 4, https://links.lww.com/CCX/B666). Comparing mode of administration (IV vs. oral) within select classes of antibiotics, we found that oral use was generally associated with lower diversity, although not statistically significant (Supplemental Fig. 5, https://links.lww.com/CCX/B666).

Gut Colonization Status on ICU Admission and Lower Alpha Diversity

Gut colonization status is not immediately available at ICU admission but can be ascertained within 3 days. Overall, 33% of participants had positive rectal swabs for MDRO colonization with VRE, CRE, or ESBL-GN at the time of ICU admission. The presence of all MDROs was associated with lower diversity, but this was only statistically significant for VRE (Table 3 and Supplemental Fig. 6, https://links.lww.com/CCX/B666).

TABLE 3.

Relationship Between Gut Colonization at ICU Admission and Alpha Diversity at ICU Admission

Gut Colonization With Multidrug-Resistant Organism on ICU Admission	Alpha Diversity			p a
	Tertile 1, Low (n = 30)	Tertile 2, Medium (n = 30)	Tertile 3, High (n = 30)
Vancomycin-resistant Enterococcus	10 (33)	6 (20)	2 (7)	0.04
Carbapenem-resistant Enterobacteriaceae	1 (3)	2 (7)	0	0.77
Extended-spectrum beta-lactamase-producing Gram-negative	2 (7)	6 (20)	3 (10)	0.37

^aχ² test was used for categorical measures, Fisher exact test was used for categorical measures with low (< 5) expected counts, and the Kruskal-Wallis rank-sum test was used for continuous measures.

Data are presented as n (%) of patients.

Regression Model for the Outcome of Lower Tertile of Alpha Diversity

Last, to identify a parsimonious list of the independent predictors of lower diversity, an ordinal logistic regression model was constructed (Table 4). The most important predictors of lower Shannon diversity at ICU admission were location before ICU admission, any culture-proven infection within the year before ICU admission, number of antibiotics used during the 3 months preceding ICU admission, and the presence of VRE on rectal swab. The final model did not include the healthcare exposure category because it was collinear with the antibiotics category. Location before ICU admission and the presence of VRE colonization on ICU admission had the highest odds ratio and were included in the final model. Prior culture-proven infection was selected as the third variable as its inclusion in the final model saw a better model fit (AIC, 184.2) than a model with number of antibiotics (AIC, 190.9). When alpha diversity at ICU admission was reexamined as a continuous measure, the same three variables were sufficient to describe a quasi-linear trend in alpha diversity (Fig. 2). To determine whether collecting a rectal swab at admission was justified, a model without VRE colonization status was also constructed but had a reduced model fit (AIC, 189.6; Supplemental Fig. 7, https://links.lww.com/CCX/B666).

TABLE 4.

Ordinal Logistic Regression Model for Outcome of Lower Tertile of Alpha Diversity at ICU Admission

Risk Factors	Unadjusted OR (95% CI)	Final Model Adjusted OR (95% CI)
Clinical characteristics
Location before ICU admission
Emergency department	Reference	Reference
Another ICU	3.06 (1.08–8.69)	3.54 (1.21–10.3)
Hospital ward	6.74 (2.19–20.8)	6.09 (1.89–19.6)
Charlson comorbidity score		—
Charlson 0–2	Reference
Charlson 3–5	5.92 (2.24–15.6)
Charlson ≥ 6	1.95 (0.72–5.27)
Mechanical ventilation	0.48 (0.19–1.22)	—
Culture-proven infections, within the year before ICU admission
Any culture-proven infection	2.87 (1.23–6.64)	2.46 (1.02–5.96)
Any multidrug-resistant organism infection	2.37 (1.07–5.27)	—
Healthcare exposure, within the year before ICU admission
Number of hospitalizations		—
0–1 hospitalizations	Reference
2 hospitalizations	0.54 (0.20–1.52)
≥ 3 hospitalizations	1.42 (0.60–3.38)
Total days in the hospital		—
0–5 d in the hospital	Reference
6–33 d in the hospital	1.93 (0.73–5.08)
≥ 34 d in the hospital	2.70 (1.05–6.99)
Antibiotics within the prior 3 mo
Number of unique antibiotics		—
0–2 antibiotics	Reference
3–5 antibiotics	3.41 (1.39–8.38)
≥ 6 antibiotics	3.59 (1.23–10.5)
Unique antibiotic classes		—
0–2 classes	Reference
3–4 classes	3.69 (1.45–9.42)
≥ 5 classes	3.21 (1.19–8.62)
Unique days of antibiotic use		—
0–1 d	Reference
2–13 d	2.98 (1.13–7.87)
≥ 14 d	2.64 (1.03–6.73)
Gut colonization with vancomycin-resistant Enterococcus at ICU admission
Absent	Reference	Reference
Present	3.62 (1.33–9.84)	4.18 (1.44–12.1)

OR = odds ratio.

Figure 2.

Shannon diversity at the time of ICU admission, stratified based on three variables. The three variables include: 1) location before ICU admission, 2) culture-proven infection within the past year, and 3) current vancomycin-resistant Enterococcus (VRE) colonization status. Points were assigned based on relative magnitude of the adjusted odds ratios and were awarded as follows: one point if the location before ICU admission was another ICU and two points if the location before ICU admission was the hospital ward; an additional point if there was an infection within the year before ICU admission; and another additional point if VRE was present at the time of ICU admission. p values were calculated using Kruskal-Wallis rank-sum test for continuous measures and Dwass-Steel-Critchlow-Fligner test for post hoc pairwise comparisons of continuous measures.

DISCUSSION

This reanalysis of RCT data identified clinical predictors for low gut microbiome Shannon alpha diversity at the time of admission to the medical ICU. The top three predictor variables were: location before ICU admission (ED vs. hospital ward vs. transfer from another ICU), any culture-proven infection within the year before ICU admission, and the presence of gut VRE colonization based on rectal swab. With these three variables, we were able to construct a parsimonious model that was associated with Shannon alpha diversity in a semi-linear manner. These results provide a novel understanding of the effect sizes for clinical predictors of gut microbiome alpha diversity at ICU admission.

Of these three top variables, two of them—location and prior infection—can immediately be gathered by taking a history or via the EMR. The third variable, VRE gut colonization status, differs from the others because it requires a dedicated test (a rectal swab) and takes 24 hours to result. Is collecting a VRE swab worth the delay and inconvenience? VRE colonization status does result in a superior model, but the other variables perform relatively well even when VRE status is unknown. If simplicity is valued or if it is important to prognosticate alpha diversity immediately, the VRE swab could reasonably be omitted. If a superior model is valued, obtaining a VRE swab is probably worthwhile. Ultimately, the answer to this question may depend on what intervention is being trialed and the planned timeframe for deploying the intervention.

This study examined additional predictors of Shannon alpha diversity at ICU admission: demographics, medical history, prior hospitalizations, antibiotic use, and MDRO colonization status. Most of these predictors demonstrated some degree of association with alpha diversity, which parallel prior studies (10, 11, 13, 14, 18, 28–30). Notably, we considered VRE gut colonization as an exposure rather than an outcome. Realistically, it takes considerable time to extract DNA from stool samples or rectal swabs for sequencing, obtain sequencing results, and analyze them to compute diversity. Conversely, a VRE rectal swab is inexpensive and rapid. VRE status may be more appropriate as a microbiome predictor than as an outcome measure.

The clinical environment immediately before ICU admission (i.e., ED, hospital ward, or another ICU) emerged as one of the strongest predictors of baseline alpha diversity. Hospitalized patients have a lower gut microbiome alpha diversity compared with nonhospitalized people due to antibiotics and hospital-associated illnesses, and these compositional changes can occur with stays as short as 4 days (30, 31). Comparatively, hospitalization within the year before ICU admission was less strongly associated with alpha diversity, suggesting that more recent events are of greater importance for diversity than events that happened in the distant past. A longitudinal study of ICU patients reported that changes in the gut microbiome during hospitalization reverted back toward baseline by a follow-up visit at a median of 71 days post-discharge (31).

We hypothesized that recent antibiotic use would be a major contributor to baseline alpha diversity (18). We found that the number of antibiotics and the number of antibiotic classes used were stronger determinants of alpha diversity than the specific antibiotics used (18, 32). Interestingly, antibiotic use was excluded from our final reduced model. Although the number of antibiotics used within the 3 months preceding ICU admission was identified as a strong predictor of diversity, our model with prior culture-proven infections performed similarly to a model with the number of antibiotics used or with both prior infections and number of antibiotics used. A likely explanation for this surprising result is that recent hospitalization and prior infections inherently reflect antibiotic exposure because antibiotics are always given for proven infections. Antibiotic use is so collinear with these other variables that it has little additional independent impact on the model. Ascertaining previous infections through prior positive microbial cultures might be clinically easier than determining the number of antibiotics taken, so we selected prior infection as our third variable over antibiotic use. Antibiotics are likely to be the single most important modifiable factor affecting the gut microbiome and our other variables of hospitalization and infection are, in a sense, functioning as surrogates for antibiotics.

This study has strengths and limitations. It is one of the first clinical predictive models for gut microbiome alpha diversity at ICU admission, and it used rigorously gathered data from a randomized controlled trial. While prior work has focused on select variables and their association with alpha diversity, our study interrogated multiple potential risk factors. To increase the robustness of the approach, we examined risk factors in many ways (e.g., antibiotic days vs. number of antibiotic classes received). We also looked at the outcome of alpha diversity as both an ordinal discrete measure and as a continuous measure. Notably, no study as of yet has defined a cutoff for an optimal or target alpha diversity in the ICU. The predictor variables that we focused on all have an established relationship with the microbiome, but our study was the first to incorporate them into a preliminary parsimonious predictive model for rapid determination of baseline gut microbiome alpha diversity at ICU admission. As far as limitations, we had a relatively small sample size. While most variables were collected prospectively during the RCT, some were collected retrospectively via chart review and may have been misclassified if patients received care that was not documented in our EMR. Although the original trial required sepsis as the ICU admission diagnosis, the source for sepsis was not specified and patients may have had different kinds of infections (lung, urine, gut). Patients were enrolled under narrow trial entry criteria, which may have led to selection bias and may affect generalizability. The model was both developed and evaluated within the same dataset, which may have resulted in an overly optimistic model performance. Before any clinical application, our findings would need to be externally validated with a second independent cohort—that is a task for future studies. Last, our underlying assumption was that patients with low baseline gut microbiome alpha diversity are those that would benefit most from future trials and therapies. However, it is unknown whether baseline diversity is a key determinant of the effectiveness of these interventions, and this assumption still requires prospective validation. Furthermore, alpha diversity examines only one aspect of gut microbiome diversity in isolation and may not adequately capture microbial function or host-microbiome interactions. Ultimately, metrics based on gut microbiome function may prove superior to alpha diversity as surrogates for gut microbiome health.

In sum, we were able to build a predictive model for gut microbiome alpha diversity at ICU admission that used only three readily obtained clinical variables. Although validation is needed before clinical implementation, these results are the first exploration into using a simple model with readily ascertained clinical variables to determine a patient’s gut microbiome alpha diversity on ICU admission.

Supplementary Material

cc9-8-e1455-s001.pdf ^{(866.9KB, pdf)}