Candida and the Gram-positive trio: testing the vibe in the ICU patient microbiome using structural equation modelling of literature derived data

The literature search uses systematic reviews of several infection prevention interventions found in The Cochrane database of systematic reviews as a starting point. These systematic reviews were identified by searching the Cochrane systematic review database using the following search terms; mechanical ventilation, ventilator associated pneumonia, blood stream infections and Candida to identify relevant systematic reviews of infection prevention interventions [22‐36]. The basic inclusion criterion for individual studies identified within these systematic reviews was patient groups requiring prolonged (> 24 h) ICU stay within studies of ICU infection prevention interventions applicable to patients receiving mechanical ventilation (MV) with, as an additional inclusion criterion, group level Candida, Staphylococcus aureus, CNS and Enterococcal infection data reported. The intervention studies were classified into four categories based on the principal component of the intervention. Studies meeting these criteria but without ICU infection prevention interventions (observational studies) were sourced to provide benchmark incidence data.

Most of the studies had been cited within either one of the systematic reviews of The Cochrane review database or within additional systematic reviews found by snowball sampling using the ‘Related articles’ function within Google Scholar [37‐59]. The snowball sampling also identified additional eligible studies. The study decant used here is as described previously [21] and is detailed in Fig. 2.

In these GSEM models, the Candida and Gram-positive bacterial infection data serve as the measurement components, the group level exposure parameters serve as the structural components and colonization with Candida, and individual Gram-positive bacteria, each represented as latent variables, link the structural and measurement components.

The incidences of VAP with Staphylococcus aureus as well as the incidences of bacteremia with each of Staphylococcus aureus, CNS and Enterococci were extracted. As Candida is generally not counted as a cause of VAP, the count of Candida as a respiratory tract (RT Candida) isolate among patients with suspected VAP was recorded along with candidemia counts. Likewise, Enterococci and CNS are rarely recorded among VAP isolates. The use of Center for Disease control (CDC) criteria, being the requirement for at least two positive cultures for diagnosis of CNS bacteremia, was recorded. Counts for all subspecies of Candida, CNS and Enterococci were included. These were each expressed as a proportion using the number of patients with prolonged (> 24 h) ICU stay as the denominator. Note that colonization with Candida, Staphylococcus aureus, CNS and Enterococcal colonization are each derived within the models as latent variables and any colonisation data within any study was not used.

The following data were used to form the structural components of the models; year of study publication, origin from trauma ICU’s, whether more than 90% of patients of the group received more than 24 h of MV, and the mean (or median) length of ICU stay (LOS) for the group. In the extraction of MV percentages, if this was not stated for any group, the percentage receiving MV was assumed to be less than 90%. In the extraction of LOS data from the studies, surrogate measures including mean (or median) length of mechanical ventilation were taken if the length of LOS was not available.

Also, the presence of any of the following group wide risk factors for candidemia and invasive Candida infection were noted; liver transplantation or liver failure, use of parenteral nutrition, surgery for intestinal perforation, pancreatitis and being colonized with Candida, however that was defined. An anti-septic exposure included use of agents such as chlorhexidine, povidone-iodine and iseganan. All anti-septic exposures were included regardless of whether the application was to the oropharynx, by tooth-brushing or by body-wash.

Topical antibiotic prophylaxis (TAP) is defined here as the group wide application of topical antibiotic prophylaxis to the oropharynx or stomach without regard to the specific antibiotic constituents. The antibiotic-based regimens often use in addition protocolized parenteral antibiotic prophylaxis (PPAP), being the protocol driven group wide use of any parenteral antibiotic used on a prophylactic basis. Group wide exposure to anti-fungal prophylaxis was identified whether this was as a single agent (SAF) or used in combination with TAP within an SDD regimen, without regard to the specific anti-fungal agent.

Scatter plots of the Candida, Staphylococcus aureus, CNS and Enterococcal infection incidence proportion data versus group mean LOS were generated to facilitate a visual survey of the entire data used in the analysis. To facilitate this visual survey, benchmarks, being the linear regression of logit transformed incidence proportion versus LOS derived for each isolate type was generated using the groups of the observational studies.

Of the 283 studies identified by the search, 135 were sourced from 37 systematic reviews (Table 1) [22‐58] with 148 found during previous searches or by snowball sampling (Fig. 2). Most studies were published between 1990 and 2010 and most had a group mean LOS exceeding ten days. A minority originated from either North American or trauma ICU’s. Twelve studies had more than one type of intervention groups and ten studies had no control group. The majority of groups from studies of infection prevention interventions had less than 150 patients per group versus more than 150 patients in the observational studies.

Of the 460 groups from 283 studies, there were 25 groups from 13 studies with mean LOS less than 5 days including the largest of which (> 120,000 patients), being a study of targeted versus universal decontamination versus standard care [61].

There was a broad range of infection prevention exposures. The majority of data for anti-fungal exposures occurred in combination with TAP exposure within Antibiotic studies in the context of an SDD regimen for which the antifungal was topical amphotericin being used in 50 groups. Exposure to anti-fungal prophylaxis as a single agent occurred (within Anti-fungal studies) in only nine groups of studies of which five studies selected patients on the basis of risk factors for invasive candida infection. The TAP exposures included either topical polymyxin or a topical aminoglycoside or both in every case except four intervention groups. PPAP, most commonly a cephalosporin, was used within ten control groups and 48 intervention groups of TAP studies.

Across all intervention categories, the incidences for Candidemia (Fig. 3) and RT Candida (Additional file 1: Fig. S1), S aureus VAP (Fig. 4) and bacteremia (Fig. 5) and CNS (Fig. 6) and Enterococcal bacteremia (Fig. 7) varied in each case by > 100 fold and ranging approximately tenfold above and below the respective benchmarks. In each case, the incidence proportions mostly straddle the respective benchmarks with exceptions noted in the figure legends.

The mean control and intervention group incidences of VAP and bacteremia for each of S aureus (Fig. 2), CNS (Fig. 3) and Enterococci (Fig. 4) were generally similar to the benchmark derived from observational groups with the exception that S aureus VAP incidences among the control groups of Antibiotic studies were generally approximately five percentage points above the respective benchmark and the mean incidences of infection for each of CNS bacteremia (Fig. 3) and Enterococcal bacteremia (Fig. 4) among the control and intervention groups of Antibiotic studies were generally approximately two percentage points above the respective benchmarks.

Four candidate GSEM models were evaluated for fit and parsimony (see Table 2; Fig. 8; Additional file 1: Figs. S2–S4). The optimal model, judged by AIC criteria, included an interaction term between the latent terms representing Candida colonization with each of the three Gram-positive bacteria colonization latent variables (Fig. 8). The size and statistical significance of this interaction term was similar in magnitude in each case. CRF predicted Candida colonization, as expected, but not bacterial colonization (Models 1–3; Additional file 1: Figs. S2–S4).

In the optimal model (model 4; Table 2; Fig. 8), the following exposures; TAP, trauma ICU admission, LOS and the interaction term with Candida colonization, displayed the strong associations with S aureus colonization. Exposure to TAP displayed positive associations with CNS colonization and Enterococcal colonization but a negative association with S aureus colonization (Table 2). The magnitude of these associations was similar in each case to that with the Candida colonization interaction term.

In all models, exposure to candidemia risk factors and TAP interventions displayed strong and consistent positive associations with Candida colonization and exposure to anti-septic, and antifungal interventions displayed strong and consistent negative associations with Candida colonization (Table 2).

There are six key limitations to this analysis. Firstly, this analysis is a group level modelling of four latent variables being colonization with each of Candida, Staphylococcus aureus, CNS and Enterococci. These latent variables and the coefficients derived in the GSEM models are indicative and intended for internal reference only. They have no counterpart at the level of any one patient or study and cannot be directly measured. Specifically, colonization data within the studies was ignored in this analysis as the mere presence of colonizing Candida may not reflect the potential biological activity towards interactions with colonizing bacteria.

The GSEM analysis takes a structural rather than statistical approach to the question of any interactions between Candida and Gram-positive bacteria. The structural approach means that a limited number of conceptually key group level factors were entered mostly as simple binary variables into intentionally simplistic GSEM models. There was no ability nor purpose to adjust for the underlying patient level risk. The true relationships between exposures and outcomes will likely be center specific, complex, graded and with multiple expoure interactions. A conventional approach to estimating exposure effects requires analytic methods such as meta-analysis which, being based on an assumption of exchangability between control and intervention groups that randomized assignement of exposures provides, allows more precise effect size estimates for specific individual interventions under study. However, assumptions that outcomes in the control groups are not influenced by infection prevention interventions within an ICU setting are questionable [74]. Moreover, a randomized assignement of individual exposure to Candida colonization is neither an ethical nor a practical intervention outside of a natural experiement resulting from group level exposures to various ICU infection prevention interventions.

The second limitation is that there was considerable heterogeneity in the interventions, populations, and study designs among the studies here as the inclusion criteria for the various studies have been intentionally broadly specified. This breadth is both a strength, in that the breadth of the group exposures is the basis for the natural experiment here which serves as the basis for the research question central within the model. It is also a limitation, in that the associations for a group wide exposure may not equate to associations at the level of an individual patient exposure.

Thirdly, several assumptions have been made for studies that failed to report key exposure and outcome variables in the analysis. Missing LOS data and percent receiving MV have been imputed and there is missing infection count data which has not been imputed. The data is provided in sufficient detail in Additional file 1 to enable replication of the analysis.

Fourth, there are a large number of studies not included here because the required infection count data was not reported. However, the differences between control and intervention group mean infection incidences noted here in the scatter plots (Figs. 3, 4, 5, 6, 7, Additional file 1: Fig. S1) are similar in magnitude and direction to the summary effect sizes for each of the three broad categories of TAP, anti-septic and non-decontamination methods, against both overall VAP and against overall bacteremia which in turn are similar to prior published effect estimates sizes seen in systematic reviews of these interventions from which most of the studies examined here were derived [19‐41].

Fifth, the various regimens of TAP, anti-septic and anti-fungal intervention that have been used within the various studies have been considered as similar within each category. This is a deliberate simplification as some, for example the anti-fungal regimens, target different body sites. Also, the duration of application of each regimen varied among the studies. On the other hand, a strength of this analysis is that it attempts to unpack the separate associations between the infection incidences and exposure to the various SDD components (TAP, PPAP, anti-fungal).

Finally, the literature search has been opportunistic rather than systematic. By using existing systematic reviews as a starting point, the key interventions can be readily identified and classified. As a consequence, the included studies have been predominately undertaken within first world country ICU’s. It is uncertain how representative this is of the microbiome for elsewhere in the world. There is some evidence that the bacteria that cause VAP vary in different parts of the world [75‐77].