Structural Equation Modelling as a Proof-of-Concept Tool for Mediation Mechanisms Between Topical Antibiotic Prophylaxis and Six Types of Blood Stream Infection Among ICU Patients

Abstract

Whether exposing the microbiome to antibiotics decreases or increases the risk of blood stream infection with Pseudomonas aeruginosa, Staphylococcus aureus, Acinetobacter, and Candida among ICU patients, and how this altered risk might be mediated, are critical research questions. Addressing these questions through the direct study of specific constituents within the microbiome would be difficult. An alternative tool for addressing these research questions is structural equation modelling (SEM). SEM enables competing theoretical causation networks to be tested 'en bloc' by confrontation with data derived from the literature. These causation models have three conceptual steps: exposure to specific antimicrobials are the key drivers, clinically relevant infection end points are the measurable observables, and the activity of key microbiome constituents on microbial invasion serve as mediators. These mediators, whether serving to promote, to impede, or neither, are typically unobservable and appear as latent variables in each model. SEM methods enable comparisons through confronting the three competing models, each versus clinically derived data with the various exposures, such as topical or parenteral antibiotic prophylaxis, factorized in each model. Candida colonization, represented as a latent variable, and concurrency are consistent promoters of all types of blood stream infection, and emerge as harmful mediators.

Keywords: Acinetobacter; Candida; Pseudomonas aeruginosa; Staphylococcus aureus; bacteremia; gut/blood microbiome; intensive care; structural equation modelling.

Figure 1

Three competing theoretical models of how exposing the microbiome (bacterial and Candida colonization) to topical antibiotic prophylaxis impacts the risk of blood stream and other infections. (a) Control of gut overgrowth (COGO), (b) colonization resistance, and (c) colonization susceptibility models. ‘Concurrency’ refers to the control and intervention groups concurrent within the same ICU. Bacterial and Candida colonization, being not easily measurable, are represented in the models as latent variables.

Figure 2

Three competing theoretical models of topical antibiotic prophylaxis mediating bacterial colonization causing blood stream and other infections are incorporated step by step into a sequence of structural equation models. Bacterial and Candida colonization are not easily measurable and are represented in the models as latent variables (ovals). The broken red arrows are the key defining steps for the COGO, colonization resistance, and colonization susceptibility models, respectively. ‘Concurrency’ refers to the control and intervention groups concurrent within the same ICU. Unbroken arrows are common to all models. Patient type refers to ICUs that have selective patient entry (e.g., trauma). Candida is not a recognized cause of ventilator-associated pneumonia (VAP), and hence here the Candida isolates are counted among respiratory tract (RT) isolates. LOS is group mean length of ICU stay. Model selection is based on Akaike information criteria (AIC).

Figure 3

Scatter plots (logit scale) and 95% CI of respiratory tract (RT) Candida incidence (a) and candidemia (b) in component (control and intervention) groups of studies of various methods of infection prevention and observational studies in the ICU. Data from 289 studies as listed in reference [87]. The mean proportion (and 95% CI) derived by random-effect meta-analysis for each category of component (observational [Ob], control [_C], and intervention [_I]) group derived from observational [Ob], non-decontamination (non-D), antiseptic (a_s), topical antibiotic prophylaxis (tap), and single antifungal (SAF) studies, is displayed. The benchmark incidence in each plot is the summary mean derived from the observation studies (central vertical line). The group-wide presence of candidemia risk factors (CRF) is identified by solid symbols versus not (open). The data in the figure are listed in reference [88].

Figure 4

Scatter plots (logit scale) and 95% CI of Pseudomonas VAP incidence (a) and Pseudomonas bacteremia (b) in component (control and intervention) groups of various methods of infection prevention in the ICU. The benchmark incidence in each plot is the summary mean derived from the observation studies (central vertical line). Abbreviations as for Figure 3. The data in the figure are listed in reference [88].

Figure 5

A model of COGO as a GSEM. Candida_col and Pseudomonas_col (ovals) are latent variables representing Candida and Pseudomonas colonization, respectively. The variables in rectangles are binary predictor variables representing the group-level exposure to the following: a trauma ICU setting (trauma50), mean or median length of ICU stay >7 days (los7), exposure to a topical antiseptic-based prevention method (a_S), exposure to a TAP-based prevention method (tap), exposure to a non-decontamination-based prevention method (non-D), use of mechanical ventilation more for than 90% of the group (mvp90) or exposure to PPAP (ppap), and exposure to azole/nystatin of amphotericin as antifungal prophylaxis. Groups with patient selection based on candidemia risk factors are factored (crf). The circles contain error terms (ε) associated with the latent variables. The three-part boxes represent the count data for Candida and Pseudomonas VAP (v_can_n, v_ps_n) and bacteremia (b_can_n, b_ps_n), each of which is logit-transformed with the total number of patients in each group as the denominator, using the logit link function in the generalized model of the GSEM. The Akaike information criterion (AIC) is 3974. The figure is adapted from reference [88] and used here under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/) (accessed on 12 November 2024).

Figure 6

A model of colonization resistance as a GSEM. The model is as for Figure 5 but includes concurrency (CC) with a group exposed to TAP as a factor. The AIC is 3928. The figure is adapted from reference [88] and used here under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/) (accessed on 12 November 2024).

Figure 7

A model of colonization susceptibility as a GSEM. The model is as for Figure 6 but includes an interaction between the latent variables representing Candida and Pseudomonas colonization. The AIC is 3921. The figure is adapted from reference [88] and used here under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/) (accessed on 12 November 2024).