Hospital Context Determinants of Variability in Healthcare-Associated Infections Prevalence: A Multi-Level Analysis

1. Introduction

Healthcare-associated infections (HAI) represent one of the major challenges in patient safety, and are associated with increased mortality, length of stay, morbidity and health expenditure. In Europe alone, they are estimated to represent almost 2.5 million disability-adjusted life years [1,2]. In the last point prevalence survey (PPS) of HAI and antimicrobial use in European acute care hospitals by the European Centre for Disease Prevention and Control (ECDC), in 2023, Portugal had a HAI prevalence of 9.9%, after validation. The result was above the ECDC projection, when adjusted for patient case-mix and hospital characteristics [3].

Extensive research has demonstrated how compliance with hand hygiene [4] and the implementation of preventive strategies in a bundle approach [5,6,7,8] may help decrease the risk of these infections. However, the role of structural variables – both infrastructure and human resources – and process variables, namely related to surveillance and infection prevention programs, remains to be established, and has seldom been researched. Published systematic reviews suggest that higher nurse staffing and the establishment of surveillance for over 5 years may be associated with decreased incidence of hospital-acquired pneumonia in intensive care units (ICU) and of surgical site infection (SSI) after colorectal surgery, respectively [9,10,11].

Better outcomes require improved processes and adequate resources [12]. Therefore, understanding the relationship between context and HAI is essential to optimize patient safety at the hospital setting. At the same, HAI are heterogeneous and affect services disproportionally, not only in terms of their overall incidence, but also in terms of the relative frequency of specific HAI [3]. Thus, it is relevant to understand how structure and process indicators relate to the different HAI in each major department, to implement maximally effective targeted interventions. One size may not fit all.

As such, this study aimed to assess whether HAI prevalence varies between hospitals, ICU, surgical and medical departments, and whether such variance may be explained by hospital characteristics, resources or implemented multimodal strategies.

2. Methods

2.1. Study Protocol

A cross-sectional study was conducted in Portuguese hospitals participating in the PPS of the ECDC, following their protocol [13], and under the supervision of the Portuguese Program for Infection and Antimicrobial Resistance Prevention of the Directorate-General of Health. Healthcare professionals collected data from patient medical records and nurses’ notes and registered them using the HelicsWin software application. All data collection in each ward was performed on one single day. Data on all hospitals were collected between May 15th and May 19th, 2023, with the exception of one hospital, that collected data on May 30th, 2023. Patients younger than 18 years of age were excluded from this study. One hospital was excluded as it did not provide data on structural and process variables.

2.2. Study Variables

Data were collected at hospital and patient levels. Data at the individual patient level included patient’s age, sex, McCabe score, which assesses the severity of patients’ comorbidities (ultimately fatal, rapidly fatal non-fatal, or unknown), and device use (central vascular catheter, intubation and/or urinary catheter).

Structural variables included the total number of acute beds, categorized in < 250 beds, 250-500 beds, and >500 beds, total number of isolation rooms in the hospital, hospital type (primary, secondary, specialized and tertiary), and location by health administrative region, divided into Norte, Lisboa e Vale do Tejo (LVT), Centro and Other, the latter encompassing the hospitals from Algarve, Alentejo, Madeira and Azores. The ratio of doctors and nurses full time equivalent (FTE) working in infection control per hospital bed, number of professionals working on antimicrobial stewardship per hospital bed and existence or not of an infection prevention and control plan and report approved by the hospital administration were also included.

Process variables included the use of masks, either on routine care or universally, hospital participation in surveillance networks for SSI, HAI in ICU, Clostridioides difficile infection, antimicrobial consumption, antimicrobial resistance, availability of clinical tests on weekends and availability of screening tests on weekends.

Furthermore, data on multimodal strategy use for implementation of infection prevention and control (IPC) interventions were collected, following the World Health Organization’s Core component 5 of infection prevention [14]. Having an IPC multidisciplinary team, regular link to develop multimodal strategies with colleagues, and the use of bundles or checklists were available in binary responses (yes or no). System change, education and training, monitoring and feedback, communication and reminders, and safety climate and culture change were each assessed with three possible answers: element not included in multimodal strategies, element included to a degree, and element fully included. The exact definitions are available on the point prevalent survey protocol [13] and in supplementary Table S1.

2.3. Statistical Analysis

Each observation corresponded to one patient. Descriptive data are reported as absolute and relative frequency and medians with interquartile range, where applicable. Analysis considered patients clustered in hospitals, in a two-level hierarchical data structure. Variance of the outcome was given by τ². An initial model was built for the entire study population, using a multi-level logistic regression, by fitting fully unconditional random-effects models with random intercepts at the hospital level. This model was used to assess the variance of HAI between hospitals. Only catheter-line associated bloodstream infections (CLABSI), SSI, catheter-associated urinary infections (CAUTI) and intubation-associated pneumonia (IAP) were considered in the outcome, since these represent not only the more prevalent and relevant infections, but also those who are under local and national surveillance, following the HAI-Net protocol [3,15]. The codes from the ECDC’s PPS used to define these infections are available as supplementary Table S2, and the flowchart of inclusion criteria for HAI is available as supplementary Figure 1. The null model was used to estimate the intraclass cluster coefficient (ICC), which provides the proportion of total observed individual variation in the outcome that may be attributable to between hospital variation. It is estimated as τ²/(τ²+(π²/3)), where π refers to the mathematical constant, approximately equal to 3.14159. The same model was used to estimate the median odds ratio (MOR), defined as the median value of the odds ratio (OR) between the hospital at highest risk and the hospital at lowest risk of infection, when randomly selecting two patients with the same covariates from different hospitals.

Study variables were tested in a univariate multi-level logistic regression as risk factors for infection. Variables with a p-value ≤0.2 in univariate analysis were tested in a multivariable model [16]. A two-sided significance level of 0.05 was considered. Collinearity was assessed in the multivariable model using the generalised variance inflation factor (GVIF). To make GVIF comparable across dimensions, GVIF^(1/(2*Df)), where Df represent the number of coefficients in the subset, was used [17]. A value of 5 or above was interpreted as presence of collinearity, and any variable meeting the threshold was dropped. The contextual effect of each included variable was given by its β-coefficient and respective p-value.

The analysis was repeated for three different subgroups: ICU, the surgical departments and the medical departments. For the analysis of ICU, stroke and coronary units were not included. The outcome considered the same infections as for the entire population. In surgical department, the outcome considered was SSI alone, as these are the only HAI under surveillance in those wards. Hence, admitted patients without previous surgery were excluded in this sub-analysis. In the medical department, the outcome was CAUTI, which is the most prevalent infection in these wards and the one under surveillance in that context.

The analysis was performed using R, version 4.4.0, using ‘merTools’, ‘lme4’ and ‘car’ packages.

3. Results

A total of 18 261 patients from 119 hospitals were included in the analysis, of which 736 were in 56 ICU, 3 160 in 72 surgical departments and 8 081 in 90 medical departments (Figure 1). The overall prevalence of any HAI under surveillance (SSI, CLABSI, IAP and CAUTI) was 3.5%. Baseline characteristics of the population, both at the individual and hospital levels, are presented in Table 1, for the entire population and per department. Most hospitals had fewer than 250 beds, although patients were almost evenly distributed across differently-sized hospitals. HAI prevalence was 7.9% in ICU, whereas SSI prevalence was 5.9% in surgical departments and CAUTI prevalence was 1.7% in medical departments.

The estimated variance of the random effects in the null model was 0.2579, corresponding to an ICC of 0.07 and a MOR of 1.62. The surgical and medical departments had a variance of 0.1875 and 0.1796 and a MOR of 1.51 and 1.50, respectively, both with an ICC of 0.05. The variance in ICU was markedly lower – 0.0299, corresponding to an ICC of 0.01 and a MOR of 1.18. The simulated random effects are plotted in Figure 2. The effect is presented in odds ratio, between each hospital and the overall average.

Univariate and multivariable analyses are presented in Table 2, overall and per department. A total of 15 variables were included in the multivariable multi-level logistic regression model for the entire hospital. Age and number of devices were identified as individual-level variables associated with higher risk of infection, whereas being in the 4th quartile in number of isolation rooms, being in the 4th quartile in the ratio of infection prevention doctor to number of acute beds, being in the 2nd and 3rd quartile in the ratio of infection prevention nurses to number of acute beds and use of masks, both universally and in care, were significantly associated with higher infection at the hospital-level. Having screening tests on one day at weekends, as well as having additional methods or initiatives to improve team communication across units and disciplines, when compared to use of reminders, posters or advocacy/awareness-raising to tools were significantly associated with lower infection.

In ICU, only the number of devices inserted in each patient was associated with HAI prevalence. No contextual variable included in multivariable analysis met the threshold.

In the surgical department, older age, ultimately fatal comorbidities assessed by the McCabe score, being in a specialized hospital and having more of infection prevention doctors and nurses per acute bed were significantly associated with SSI. A full investment in safety climate in the department, with empowerment of teams and individuals, when compared to no inclusion of this element in a multimodal strategy for infection prevention and control was associated with lower SSI.

In Medicine departments, age and number of devices were significantly associated with CAUTI, and a larger ratio of nurses in infection prevention per hospital acute bed (4th quartile vs 1st quartile) was associated with decreased CAUTI. In all models, no variable had to be removed due to collinearity, as GVIF^(1/(2*Df)) was always below 5 for every included variable (supplementary Table S3).

4. Discussion

This study found that structural and process variables association with HAI may be different depending on the services or department considered within hospitals, and on the specific types of HAI. In the ICU, variance is much lower and the number of devices per patient is the only variable associated with HAI prevalence. In surgical and medical departments, however, several context variables appear to influence HAI prevalence.

The apparent immunity of ICU to contextual variables may be due to the fact that ICU are more similar to each other than surgical or medical departments, which encompass a variety of different services whose distribution is affected by the hospital size and type. At the same time, the workflow in ICU relies more heavily on detailed protocols and procedures than their counterparts, which may translate in a more systematic and standardised approach to patient care, hence, with lesser variability between settings. Our data on process and structural variables does not show a difference in variability between departments. However, it may fail to grasp these nuances as most data is aggregated at the hospital level. Finally, it is worth considering that the effect of devices may be so significant, when their prevalence is higher as is the case of ICU, that it may overshadow other smaller potential determinants. These hypotheses require further research, directed to the questions at hand.

In surgical departments, the empowerment of teams and individuals as owners and protagonists of quality improvement interventions and the existence of a safety climate and psychological comfort were found to have a negative association with HAI prevalence. The visible support and commitment of managers and leaders in strengthening a culture of patient safety was not significantly associated. These results emphasize the role of quality improvement interventions in the battle against HAI. Portugal implemented such a program, between 2015 and 2018, named Stop Infeção Hospitalar!, aiming to reduce the incidence rate of the four typologies of HAI here researched – CLABSI, CAUTI, IAP and SSI – in 12 hospitals, following a previous positive experience [18], and is currently developing a similar program, called Stop Infeção Hospitalar-2.0, involving 22 acute care hospitals and aiming at a 50% percent reduction, in three years, in the incidence of the same HAI [19]. This suggests that, at least in surgical departments, the success of the project lies both on the quality of the proposed bundles and the compliance with them, and on the philosophy of local teams and professionals, their empowerment, psychological safety and cooperation and participation in the endeavour. This has been also emphasized in other studies [20]. Safety culture change was marginally insignificant in the entire population and almost similar to chance in medical department, raising the question that the effect of this strategy may be different in different services and departments, depending on the professionals and department cultures.

When considering all services from included hospitals, an investment in active communication beyond the use of common awareness-raising tools was found to be significantly associated with a smaller prevalence of HAI. Even though the implementation of the IPC assessment framework has been recently shown to be associated with a decreased prevalence of HAI [21], to the best of our knowledge, no study has underlined the individual role of communication as a main driver for better health results. However, communication strategies have been found to be associated with improved healthcare delivery results in different contexts [22,23], and our finding supports a renewed investment in this research field.

The number of isolation rooms has also been found to be positively associated with lower HAI prevalence in the entire sample, when adjusted for hospital bed size and type. This result has not been found in the sub-analyses per service and department. However, the variable relates to the number of isolation rooms in the hospital, not per service, and therefore it may have no practical meaning in the sub-analyses performed. Likewise, the availability of screening tests on weekends has been found to have a negative association with HAI prevalence at the entire sample level only. Isolation rooms have been found to reduce the rate of HAI in pediatric ICU and in the COVID-19 context [24,25], but published papers are lacking, suggesting these results should be interpreted with caution, warranting further research.

The positive association between the ratio of IPC nurses and doctors per acute bed and the prevalence of HAI, which was found in the entire sample but also on surgical and medical departments, should not be interpreted as a causative association and it should not lead to the conclusion that an investment in specialized personnel is deleterious to infection prevention efforts, but rather that an artificial increase in HAI rates may occur. These artificiality is well supported in the literature, that consistently found that better registry, active post-discharge case finding and better audits are associated with a report of higher infection rates [26,27]. This effect was not found when the participation in surveillance networks was assessed, even in specific cases, as was the case of participation in surveillance network of HAI in ICU, and participation in surveillance network of SSI in surgical departments. However, the cross-sectional nature of the paper may fail to find the parabolic effect of surveillance on HAI, which tends to artificially increase rates, and then reduce them after the 5-year mark [9]. At the same time, the high ratio of IPC personnel may reflect a prior response to the higher prevalence of infection, and represent a consequence of high HAI rates. Due to the transversal nature of the analysis, one may not clearly define the direction of the association found.

Our research also found that, although univariate analyses suggest that hospital size may have a role, albeit indirect, in HAI rates, such effect was unapparent after adjustment to other variables. Even though the paper by Zingg et al. did find a significant association, they compared larger hospitals to smaller hospitals, with a cut-off at 650 beds, which is hardly meaningful at many national settings [16].

To the best of our knowledge, this is the first paper to consider that different departments may benefit from different IPC strategies and have different determinants of their success. Few papers have used the ECDC’s PPS to assess contextual factors on HAI prevalence in the entire hospitals [6,28,29] Barbadoro et al. and Zingg et al. both used a multilevel logistic regression approach, yet the former performed bivariate analysis only. The latter found that larger, tertiary care hospitals had a positive association with HAI prevalence, as discussed, and that private-for-profit hospitals had a negative association. [16,29] Palaiopanos et al. followed a different approach and used a multivariable linear regression analysis of hospital factors associated with HAI [28]. Bed occupancy was the only variable with a significant – and positive – association. The large sample of our study, obtained by the participation of virtually every hospital in the country, together with the inclusion of multimodal strategies implementation, strengthen the internal validity of our findings. The inclusion of HAI that are currently under surveillance at each local setting makes our findings more relatable to everyday professionals in infection prevention and control, improving the practicality of interventions that address some of the highlighted issues presented.

This study has some limitations. As a cross-sectional study, it disallows researchers from drawing causality from any finding, and absence of significant association does not mean that the variable is not associated with infection. The paper addresses variance between institutions and, hence, every time hospitals share an intervention – having a multidisciplinary team – or fail systematically to do so, the variable will fail to significantly explain any variation, regardless of its real-world potential association with HAI rates. Also, some variables have inherent limitations, as is the case of the implementation of bundles or checklists, which are considered together despite being two diverse things. Besides the limitation of being unable to consider bundles and checklists separately in the analysis, it is also possible that the interpretation of multimodal strategies inclusion may have not been consistent throughout the hospitals, although random variation is not expected to affect the validity of our findings. Finally, the high proportion of unknown values in multimodal strategies implementation may have hindered the potential conclusions regarding their effect, since missing values may have biased the comparison.

5. Conclusion

Different hospital departments seem to have different infection prevention and control needs and potentially benefit from different ICP interventions and strategies. The results of this paper may help implement targeted interventions that may achieve optimal results in each department, considering the specificities of local realities.