Introduction

Respiratory tract colonization with Candida (RT Candida) among patients with suspected ventilator-associated pneumonia (VAP) has been reported in numerous studies [1–113]. Both the overall incidence and the clinical significance for the individual patient are uncertain. True Candida pneumonia in this patient group is thought to be rare [114, 115]. Among a tally of 2,490 isolates from 24 studies, fungi (species unspecified) accounted for only 0.9 % of pathogenic isolates [116]. While current guidelines [114, 115] do not recommend treatment of RT Candida, it remains of interest for at least four reasons.

Firstly, colonization with Candida is believed to be a key intermediary step towards invasive candidiasis, although the role of RT Candida in this respect is unclear. The respiratory tract, being a site not normally colonized by Candida, may provide a unique insight into factors influencing the incidence of Candida colonization. Thirdly, RT Candida may be a risk factor for specific bacterial infections due to molecular interactions [117, 118].

Finally, the influence of topical antibiotic (TA) use as a method to prevent ICU-acquired bacterial colonization and infection, as within studies of selective digestive decontamination (SDD) and selective oro-pharyngeal decontamination (SOD) [119, 120], on the incidence of Candida colonization is of longstanding interest. This question was raised in the first study of SDD [74]. There appear to be subtle contextual effects of using topical antibiotics in the ICU on the incidence of candidemia [121], as with bacteremia [122] and also as with ventilator-associated pneumonia [123], which are evident only through benchmarking the control group rates of these studies and which are not seen in studies of non-TA methods of ICU infection prevention.

RT Candida data is available from numerous studies of a broad range of VAP prevention methods which have been reviewed in systematic reviews [119, 120, 124–135]. Among this evidence base are those with concurrent versus non-concurrent study designs, together with other study designs including those without any intervention. This heterogeneous evidence base provides a natural experiment [136, 137] with which to address some of these questions at the group level, using methods as used in the analysis of cluster randomized trials.

Materials and methods

Study selection and decant of groups

The literature search and analytic approach used here is as described previously [121]. These seven steps (Fig. 1; numbered arrows) are as follows;

Fig. 1
figure 1

Search method, screening criteria, and resulting classification of eligible studies and subsequent decant of component groups. The seven numbered arrows to the right represent steps in the process as discussed in the methods; steps 1 to 5 refer to studies, and steps 6 and 7 refer to component groups decanted from the studies being control (rectangles) and intervention (ovals) groups from ICU-based studies of infection prevention methods and observation groups from cohorts of ICU patients without a prevention method under study (diamond). The horizontal dotted rectangles represent the group contrasts used toward the calculation of the contextual effects of the component groups, each versus the observation groups as the reference category. Note; the total numbers do not tally, as some systematic reviews provided studies in more than one category, and some studies provided groups in more than one category. Note; SDD includes SOD and other methods based on the use of a topical antibiotic (TA)

Full size image
  1. 1.

    An electronic search of PubMed, The Cochrane database and Google Scholar for systematic reviews containing potentially eligible studies was undertaken using the following search terms; “ventilator-associated pneumonia”, “mechanical ventilation”, “intensive care unit”, each combined with either “meta-analysis” or “systematic review” up to December 2013.

  2. 2.

    Systematic reviews of studies of patient populations requiring prolonged (>24 hours) ICU admission were then streamed into one of three categories; systematic reviews containing studies in which there was no intervention, studies with SDD as the intervention, or studies with an intervention other than SDD, for the prevention of VAP. For the purpose of this study, SDD is defined here as the use of protocolized topical antibiotic prophylaxis applied by the gastric or oro-pharyngeal route in the intervention group, with or without the additional use of a parenteral antibiotic or any anti-fungal agent.

  3. 3.

    The studies were screened against the following eligibility criteria. Inclusion criteria; incidence data for ventilator-associated pneumonia extractable as an incidence proportion being expressed as a proportion of numbers of patients among patients with an ICU stay of at least 24 hours. Exclusion criterion; studies limited to patients with the acute respiratory distress syndrome. Studies in a language other than English were included when the required data had been abstracted in an English-language systematic review.

  4. 4.

    A hand search was undertaken for additional studies not identified within systematic reviews but otherwise meeting the eligibility criteria.

  5. 5.

    All eligible studies were then collated and any duplicate studies were removed.

  6. 6.

    Groups of patients receiving mechanical ventilation from studies without a VAP prevention method under study were labelled as observational groups. The studies of intervention studies were classified as follows. The non-TA-based methods of VAP prevention used interventions other than topical antibiotics. These were usually delivered at either the gastric, airway, or oral sites. The SDD studies were further sub-classified as to whether the control group was concurrent and co-located within the same ICU as the intervention group (concurrent control) or not (non-concurrent).

  7. 7.

    The component (control and intervention) groups were decanted from each study as follows;

    • The control and intervention groups from non-TA based methods were classified as indicated in the original study

    • Among studies of SDD, all groups that received prophylaxis with any regimen of topical antibiotic, whether or not an anti-fungal was included in the regimen, were designated as an SDD intervention group and all other groups from SDD studies were classified as a control group.

Data extraction

The RT Candida figure is the number of patients with Candida isolates from respiratory sampling per 100 patients with prolonged (>24 hours) stay in the ICU, whether or not VAP has been documented. In addition, the following were also extracted where available; the overall incidence proportion of VAP, the incidence of candidemia, the incidence of Candida colonization at non-respiratory tract sites without regard to how this had been defined in each study, and the proportion of admissions for trauma. Each of these were expressed as a proportion, using the number of patients with prolonged (>24 hours) stay in the ICU as the denominator. Other parameters extracted were whether the mode of diagnosis of VAP required bronchoscopic sampling and whether topical placebo had been used to achieve observer blinding.

Caterpillar plots

To generate caterpillar plots, the RT Candida data were logit transformed for analysis as previously [137]; with the total number of patients as the denominator (D), the number of patients with RT Candida as the numerator (N), and R being the RT Candida proportion (N/D), the logit(RT Candida) is log(N/(D−N)) and its variance is 1/(D*R*(1–R)). Note that for any group with a zero event rate (N = 0), the addition of the continuity correction (i.e., N + 0.5) is required to avoid indeterminate transformations of mean and variance. Using these pre-calculated logits and logit variances, group specific 95 % confidence intervals, summary logits and the associated summary 95 % CIs were generated using the ‘metan’ command in STATA. On the logit scale, the 95 % confidence intervals for a proportion are symmetrical and remain within the interval of 0 to 100 %.

For each category of component group the summary mean logit RT Candida and associated 95 % confidence interval were calculated using random effects methods. These were then back-transformed to the percentage scale. The benchmark is the summary mean RT Candida per 100 patients derived from the observational studies, and the benchmark range is the 95 % prediction interval.

Bivariate plots and confidence ellipses

To assess correlation of RT Candida with candidemia incidence, the logit-transformed data was assessed by two methods; a 95 % prediction ellipse [138–140], and linear regression. The prediction ellipse method enables the correlation as observed in other studies to be benchmarked. The relationship between logit-transformed RT Candida with year of publication was assessed using locally weighted regression and smoothing scatterplot (LOWESS) [141].

Statistical analysis

A regression model of RT Candida proportion was developed using GEE methods, as these accommodate any intra-cluster correlation (‘xtgee’ command in STATA; release 12.0, STATA Corp., College Station, TX, USA). In this analysis, the predictor variables were: (1) the component group membership, being either membership of a group from an observational study, a control group, or an intervention group, (2) type of intervention under study, (3) the use or non-use of topical placebo, and (4) whether the mode of diagnosis of VAP required bronchoscopic sampling. As a sensitivity analysis, the GEE regression model analysis was repeated limited to studies obtained from systematic reviews. In addition to the Poisson model, the GEE regression model was undertaken with both binomial and negative binomial models as additional sensitivity tests.

Results

Characteristics of the studies

Of the 113 studies identified by the search [1–113], 72 were sourced from 13 systematic reviews and a further 40 sourced from elsewhere (Table 1; Fig. 1). The majority of SDD studies were published in the 1990s, and all but four studies of SDD were European in origin. Two studies were supplemented with data from published doctoral theses [76, 113] or related publications [74].The studies are detailed in the Electronic Supplementary Material (ESM) file.

Table 1 Characteristics of studies a
Full size table

A total of 191 component groups were decanted from these 113 studies, with 36 groups from observational studies (ESM file Table S1), 71 groups from studies of various non-TA methods of VAP prevention (ESM file Table S2), and 84 groups from studies of SDD having either a non-concurrent (ESM file Table S3) or concurrent design (ESM file Table S4). Eleven studies had more than one observational, control, or intervention group. Two studies had both concurrent and non-concurrent control groups. Group-wide risk factors for candidemia were identified in only six studies. Three SDD studies used a regimen not containing an anti-fungal [71, 111, 112].

Candida colonization

A measure of Candida colonization not limited to patients with suspected VAP and not limited to respiratory sites was reported for 32 studies including 18 of the SDD studies (Table 1). There was a wide range in this incidence (Figs. S5 & S6), with the incidence among concurrent control (p = 0.066; Table S5) groups from studies of SDD being higher versus the incidence in the observational groups.

The incidence of Candida colonisation, at sites other than the respiratory tract and not restricted to patients with suspected VAP, together with candidemia incidence, were available from 40 groups from studies of all types. There were too few groups to discern a significant relationship between these incidences or to generate a robust prediction ellipse using the non-TA studies (Fig. S5).

RT Candida incidence

The mean RT Candida incidence among the 36 observational groups is 1.3 (95 % confidence interval; 0.9–1.8 %) (ESM file Fig. S1). This is the RT Candida benchmark. There was no significant trend in RT Candida incidence by year of study publication and a LOWESS line is presented (Fig. 2). Twenty-three of the 32 control groups of the concurrent control design studies were above this LOWESS line. The mean RT Candida incidence among the control groups was significantly higher than in the intervention groups from concurrent design SDD studies (p = 0.001; Fig. S4).

Fig. 2
figure 2

RT Candida incidence versus year of publication. Scatter plot of RT Candida incidence in observational groups (open black circles) versus year of study publication for which the linear regression was non-significant (not shown, p =0.72) and hence a LOWESS regression line is given. Also shown are the control groups from studies of non-TA methods (open blue triangles) and also concurrent control (CC) studies of SDD/SOD (closed red triangles). Note that the symbols are proportional to group size and the y axis is a logit scale

Full size image

The RT Candida incidence was highest amongst the concurrent control groups of the SDD studies versus other types of component group (Fig. 3, ESM file Figs. S1–S4). The effect of membership of the various categories of component group on RT Candida was examined in GEE models of RT Candida together other group level variables (Table 2). The effect of membership of a concurrent control group of an SDD study was significant (p = 0.021). The group wide presence of candidemia risk factors (p = 0.001) and the proportion of admissions for trauma (p = 0.004) were also significant factors in the model.

Fig. 3
figure 3

Scatter plots of RT Candida incidence. Scatter plots of RT Candida incidence in component (C = control; I = intervention) groups of non-TA and non-concurrent (NCC), and concurrent control (CC) studies of SDD (antibiotic methods) of VAP prevention (blue symbols). The benchmark RT Candida is the summary mean (central vertical line) derived from the observation studies (black symbols) together with the 95 % prediction intervals (benchmark range; outer vertical lines). Studies with MV for >48 hours for <75 % and one outlier study [84] are not shown. Note that the x axis is a logit scale and that studies with a zero RT Candida can only be shown on this plot through use of the continuity correction. These data are displayed in more detail as caterpillar plots (Figs. S1–S4)

Full size image
Table 2 Logit RT Candida incidence; generalized estimating equation models (all groups) a
Full size table

Candidemia

The incidence of candidemia was reported for 101 of the component groups (Table 1, Fig. S7). The incidence among concurrent control (p = 0.024) groups from studies of SDD was higher than that in the observational groups.

Both candidemia incidence and RT Candida incidence data were available from 19 groups from either observational studies or from studies of non-TA methods (Fig. 4a). The scatterplot presenting the bivariate relationship among these 19 groups, together with the linear regression line and a 95 % prediction ellipse, is shown using logit scales for each axis (Fig. 4a). This linear regression and prediction ellipse are in turn used to benchmark the groups from the studies of SDD (Fig. 4a–c). Whilst most of the control groups from studies of SDD with a non-concurrent (Fig. 4b) and concurrent design (Fig. 4c) are within this benchmark prediction ellipse, shift to the right and upward is apparent for the control groups in the latter plot (Fig. 4c).

Fig. 4
figure 4

Candidemia incidence versus RT Candida. a Scatter plot of candidemia incidence versus RT Candida incidence in observational groups and control groups from studies of non-TA methods (open black circles) for which both data were available. Both a linear regression (p =0.057) and a 95 % prediction ellipse based on these groups are shown. Also shown are the intervention groups from studies of non-TA methods (closed triangles). Also shown are the mean (central) and 95 % prediction lines (outer) derived for the full set of benchmark groups for both RT Candida (vertical) and candidemia (horizontal). Note that both the x and the y axes are a logit scale. b Scatter plot of candidemia incidence versus RT Candida incidence in control (closed green triangles) and intervention (open green triangles) groups from studies of non-concurrent design studies of SDD (open green triangles) for which both data were available. The linear regression, 95 % prediction ellipse and mean, and 95 % prediction lines from a are shown for reference. c Scatter plot of candidemia incidence versus RT Candida incidence in control (closed green triangles) and intervention (open red triangles) groups from studies of concurrent design studies of SDD (open red triangles) for which both data were available. The linear regression, 95 % prediction ellipse and mean, and 95 % prediction lines from a are shown for reference. Note that one outlier study [84] is not shown

Full size image

Discussion

This is a meta-analysis of the incidence of respiratory tract colonization with Candida (RT Candida) among ICU patients receiving mechanical ventilation within studies in the literature. This analysis has examined the relationship between each of RT Candida and candida colonization at other sites with candidemia, as well as the relative importance of selective digestive decontamination (SDD) versus other candidate risk factors towards RT Candida. Only 45 of the studies are common to this meta-analysis and to the previous meta-analysis of candidemia [121]. This previous meta-analysis [121] was not restricted to the patient population receiving mechanical ventilation, and had included a higher proportion of studies with group-wide risk factors for candidemia (25 of the 103 studies [121]), versus only six of the 113 studies included here. As a consequence, there is a lower incidence of candidemia among the studies here versus previously [121]. However, even within this differently selected set of studies, the incidence of candidemia is again found to be higher among the control groups of SDD studies with a concurrent design than any other type of component group (Table 1).

There is a higher incidence of both Candida colonization at respiratory (RT Candida) and other sites among the control groups of SDD studies with a concurrent design versus other groups (Table 1). This higher incidence of RT Candida cannot be explained by year of publication (Fig. 2), nor in regression models that include the group-wide presence of risk factors for candidemia or proportion of trauma admissions or mode of respiratory sampling (Table 2).

There is a correlation between candidemia and RT Candida among the studies here for which data is available (Fig. 4). However, the relationship between RT Candida and candidemia is more complex than a simple linear correlation for the following reasons. Candidemia, with an incidence of approximately 1 % amongst ICU patients [121], is a rare outcome, and studies with fewer than 100 patients may have one or no cases [62, 142–147]. The relationship described here is at the group level rather than at the patient level. In relation to non-respiratory Candida colonization, the patient-level association has recently been examined in a multi-center study [148]. The relative risk for invasive candidiasis in the mechanically ventilated ICU patient population differs for throat, perineum, and urine sites of colonization, and also for different sampling time points [148]. Measures of non-respiratory Candida colonization among the studies here were poorly documented in relation to exact sites and timings, and in any case were available from less than half of the studies surveyed.

It should be noted that the clinical significance of RT Candida is unclear [149–155], and current consensus guidelines recommend against its specific treatment in the absence of either clear histological evidence for pulmonary infection, which is rare [145–147], or evidence of invasive disease [114, 115]. However, Candida colonization continues to remain of interest from both the individual and the population perspective.

At the level of the individual, the clinical significance remains unclear, with conflicting results of studies of Candida colonization of the respiratory tract among ICU patients. Some investigators have found that Candida colonization of the respiratory tract is associated with a worse outcome [149] in association with evidence of increased systemic inflammation [150]. However, a subsequent pilot study of antifungal therapy for RT Candida did not find sufficient evidence of benefit to justify proceeding to a full-scale controlled trial [151]. By contrast, other workers have not found an association with a higher mortality risk [144, 152] in ICU patients, even though there were higher disease severity scores or degree of organ dysfunction at ICU admission. Moreover, no apparent outcome benefit associated with the use of empiric systemic anti-fungal therapy in this patient group was found in either this study [152] or in a large multi-center study [153].

From the population perspective, Candida colonization is an important constituent of the ICU microbiome. RT Candida could increase the risk for co-infection with antibiotic resistant bacteria in the airway [154], through molecular interactions with bacterial pathogens [117] for which anti-fungal therapy may be protective [118].

RT Candida has potential use as a more readily available indicator for benchmarking Candida colonization incidence in the patient group receiving mechanical ventilation. RT Candida is used here for this purpose so as to benchmark Candida colonization across different studies that have examined a variety of interventions, whether using TA-based regimens such as SDD or non-TA-based methods for VAP prevention.

The effect of SDD on Candida colonization is unclear. On the one hand, SDD as a regimen comprising topical antibiotic and anti-fungal agents appears to be protective against fungal colonization, infection, and possibly even mortality [119, 120]. Indeed, the protection derived by SDD appears to outperform that obtained using azole antifungal prophylaxis in this patient group [120]. On the other hand, this protection is not apparent in the largest study, which had a non-concurrent design [155]. Moreover, SDD may have complex effects on the ICU microbiome. Indeed, in the first SDD study [74] it was asked whether this indirect effect of SDD might confound any attempt to estimate the direct effects using a conventional concurrent study design.

An uncalibrated analysis of the available Candida colonization data, whether as RT Candida or as Candida colonization not restricted to respiratory sites and not restricted to patients with suspected VAP, is consistent with what appears to be a near halving in colonization incidences, as implied in the meta-analyses of the concurrent design SDD studies [119, 120]. However, on closer scrutiny and using the RT Candida benchmark for calibration, the true impact of SDD on the incidences of Candida colonization as well as on candidemia would appear to be a near doubling amongst the concurrent control groups of SDD studies (Table 1). This occurs presumably as a result of an indirect contextual effect through inapparent cross infection [156]. By contrast, the effect of SDD on RT Candida in studies that are non-concurrent is insignificant (Table 2), as observed elsewhere [155].

It is not the intention here to examine the substantial number of different SDD regimens but rather the component groups from the two broad categories of TA and non-TA studies. In any case, it should be noted that complete Candida decolonization using SDD is difficult to achieve [157, 158]. Two recent studies of ICU patients that were colonized with Candida and were receiving SDD provided conflicting evidence that the administration of nebulised amphotericin additional to SDD might confer clinical benefit [157] versus harm [158]. Of note in both studies, the time to achieve 50 % decolonization with the addition of nebulised amphotericin to the standard SDD regime was 5 days in both studies. By contrast, among a multi-center study of 3,000 ICU patients colonized with Candida receiving routine systemic antifungal therapy with a mean ICU stay of 5 days, typically between 40 and 50 % remain colonized on ICU discharge [159].

There are four specific challenges in undertaking an analysis of RT Candida. First, the potential for transmission of Candida between control and intervention group patients in the same study renders the presumption of independence of RT Candida events untenable [160].

Second, for most SDD studies the primary end point was VAP occurrence, and RT Candida was an occasional secondary end-point. How studies with zero RT Candida events are optimally included in any analysis is important to the conclusions. Studies with zero RT Candida events should not be overlooked, as they provide potential evidence against a contextual effect. However, the majority of the SDD studies were smaller than 60 patients and, being a rare event (<2 % in most studies), a zero RT Candida event rate is unsurprising. As a consequence, the upper 95 % confidence intervals for these groups are non-trivial in caterpillar plots (Figs. S1–S4). Note that for a group of size N = 60 with zero events ,the upper 95 % confidence can be approximated by the ‘rule of three’ as 5 % (=3/N) [161].

Thirdly, to quantify a contextual effect requires a calibration to a benchmark range derived using for reference data from studies from comparable target populations. The final issue is one of validity: does RT Candida correlate with a clinically relevant and commonly reported end-point?

To deal with the first and third of these challenges, GEE-based analytic strategies have been used here to model the RT Candida of both control and intervention groups of all studies within a single analytic model as a statistical calibration (Table 2). There is an upward dispersion in RT Candida incidence among concurrent control groups from SDD studies away from this benchmark (Fig. 3). This upward dispersion is apparent in the GEE models as positive coefficients in association with membership of concurrent control groups within studies of SDD/SOD (Table 2).

To deal with the second and third issues, the continuity correction has been used to enable zero-event groups to be represented on the logit scale, which enables several types of graphical display for the purposes of a visual calibration (Figs. S1–S4; Fig. 3). Moreover, the validity issue is able to be addressed through an examination of the bivariate relationship between RT Candida and candidemia (Figs. 4a–c). The visual analysis of the bivariate relationship is aided by the use of a 95 % prediction ellipse in the plots, a method which is better suited to this purpose than linear regression [138–141]. All of these visual displays dramatically reveal that the component groups of all types, with the exception of those concurrent control groups from studies of SDD, each have a distribution similar to the observational groups from which the benchmark was derived. Strikingly, even the distribution of the intervention groups of the SDD studies are similar to those from which the benchmark was derived. This would imply that any apparent effect of SDD on Candida colonization and candidemia within concurrent control design studies is not explainable as simply a direct anti-fungal prevention effect occurring within the intervention group (Fig. S7).

There are several limitations to this study. This is an analysis at the group level, and is unable to take account of patient-specific risk factors for RT Candida. For example, the usage of empiric (non-protocolized) antifungal therapy in each study is an important unknown, as non-use may account for vulnerability to RT Candida at the individual level. However, it is unlikely that such unidentified patient-level risk factors would be able to account for the discrepancies noted here. Such a putative patient-level risk factor would need to be a consistently strong risk factor for RT Candida across all the studies and yet also be profoundly unevenly distributed, predominating in the groups of the SDD studies versus other groups within the broader evidence base examined here.

Another limitation is the imprecision associated with the diagnosis of VAP, which may lead to the potential for observer detection bias of RT Candida. That the mode of VAP diagnosis and the use of topical placebo were not significant factors in the regression model (Table 2) implies that this bias is likely to be minimal. Moreover, topical placebo use can be taken as a surrogate indicator of a study that was observer-blinded. A further limitation is the question of non-reporting of RT Candida amongst potentially eligible studies. However, the correlation between RT Candida and candidemia provides some validation, at least amongst those studies for which both data were available.

It is never possible to be certain that every relevant study has been obtained in a literature search or that the search has been truly adequate. Restricting the analysis to those studies obtained from systematic reviews attempts to provide the basis for an analysis of data derived through an independent and transparent search. That the findings of such a restricted analysis are similar to the full analysis would imply that the search has been adequate and that the number of missing studies required to alter the findings would need to be substantial.

Conclusion

The RT Candida incidence within observational groups of mechanically ventilated patients is 1.3 % (this is the RT Candida benchmark). The incidence of RT Candida and candidemia are correlated. There is insufficient information to discern how closely Candida colonization at other sites is correlated with candidemia. At the group level, the presence of candidemia risk factors, the proportion of trauma admissions, and membership of a concurrent control group within an SDD study are each risk factors for RT Candida. The apparent protection against Candida colonization from the use of SDD appears spurious, as the incidence of both RT Candida and colonization at other sites is higher among concurrent control groups of SDD studies than among observational and indeed other types of component group. These observations, as with similar observations for VAP [123], candidemia [121], and bacteremia [122] incidences among these studies, are paradoxical. Apart from major publication bias, or the effect of any major and as yet unidentified and mal-distributed patient-level risk factors for RT Candida, these profound discrepancies indicate a major contextual hazard associated with the topical antibiotic component of SDD on RT Candida within studies with concurrent controls. These increased incidences would be inapparent within individual SDD studies examined in isolation [156]. Abbreviations used in this paper are: ICU, intensive care unit; MV, mechanical ventilation ; SDD, selective digestive decontamination; VAP, ventilator-associated pneumonia; RT Candida, Candida among patients with ventilator-associated pneumonia.