Simulation models for transmission of health care-associated infection: A systematic review.

BACKGROUND: Health care-associated infections (HAIs) are a global health burden because of their significant impact on patient health and health care systems. Mechanistic simulation modeling that captures the dynamics between patients, pathogens, and the environment is increasingly being used to improve understanding of epidemiological patterns of HAIs and to facilitate decisions on infection prevention and control (IPC). The purpose of this review is to present a systematic review to establish (1) how simulation models have been used to investigate HAIs and their mitigation and (2) how these models have evolved over time, as well as identify (3) gaps in their adoption and (4) useful directions for their future development.
METHODS: The review involved a systematic search and identification of studies using system dynamics, discrete event simulation, and agent-based model to study HAIs.
RESULTS: The complexity of simulation models developed for HAIs significantly increased but heavily concentrated on transmission dynamics of methicillin-resistant Staphylococcus aureus in the hospitals of high-income countries. Neither HAIs in other health care settings, the influence of contact networks within a health care facility, nor patient sharing and referring networks across health care settings were sufficiently understood.
CONCLUSIONS: This systematic review provides a broader overview of existing simulation models in HAIs to identify the gaps and to direct and facilitate further development of appropriate models in this emerging field.

Health care–associated infections (HAIs) pose a serious risk for patients and providers because they cause increased morbidity and mortality, prolonged length of stay in health care facilities, increased prevalence of multidrug- resistant organisms, and psychological and financial burdens to patients, their families,and the health care system. The risk of HAIs is universal and pervades every health care facility, setting, and system globally. In Europe, the prevalence of HAIs was estimated at 5.5% and about 2.6 million new patients having HAIs have been identified annually. These infections accounted for an estimated 33,000 attributable deaths and 900,000 disability-adjusted life-years. In the United States, the estimated prevalence of HAIs in hospitals was between 2.9% and 3.5% in 2015. The burden is even higher in low- and middle-income countries (LMICs). A systematic review and meta-analysis reported that the pooled prevalence of overall HAIs in Southeast Asia, where most countries are middle-income, was 9.1%. The reported prevalence in Africa varies significantly: in Ghana, prevalence ranged between 3.5% and 14.4% in acute care hospitals, and in tertiary hospitals in South Africa and Ethiopia, it was 7.67% and 19.4%, respectively.6, 7, 8 Data on the impact of HAIs at the national level in LMICs, especially African countries, are scanty and fragmented, generating difficulty in assessing the true scale of the problems of HAIs. The actual figure is assumed to be higher owing to the lack of a functioning HAI surveillance system in these countries.

Historically, randomized control trials (RCTs), cohort studies, and case-control studies were commonly used methods to investigate the epidemiology of diseases and the epidemiology of HAI in particular. Additionally, researchers performed cluster RCTs or quasi-experimental studies to examine the effectiveness of various measures for infection prevention and control (IPC). However, performing large cluster RCTs across various health facilities to achieve generalizability and sufficient power to address important research questions is difficult. In addition, although quasi-experimental studies are more feasible and practical to conduct, the lack of randomization is a threat to the internal validity and limits the generalizability of the results to larger populations. Although simpler non-mechanistic modeling approaches, such as statistical models and analytical models, have also been used to evaluate IPC interventions, they cannot capture the complexity and dynamics of HAI transmission and the health care contexts in which the interventions are implemented. Therefore, a more comparable, reliable, and easy-to-use planning tool is needed to assess interventions and their impacts.

Modeling is increasingly being used to improve the understanding of epidemiological patterns of HAIs and to facilitate decisions on IPC intervention. Mechanistic simulation modeling that captures the dynamics between patients, pathogens, and the environment is particularly useful for studying complex systems such as the health care system. A simulation model can be used to understand the dynamics of HAIs and IPC and how various complexities influence these dynamics or to predict outcomes of IPC interventions. The latter can only be done credibly provided we understand the system. Simulation modeling provides a risk-free environment in which ideas on IPC strategies can be tested in a systematic manner without the time, costs, and risks associated with experiments conducted in a real-world setting. It is a valuable tool to guide the selection of the most appropriate empirical research to pursue and to examine the effects of IPC strategies, serving as a “virtual policy laboratory” for decision support by researchers, policy makers, public health officials, hospital managers and administrators, and other health care decision-makers.

Like other modeling methodologies that try to predict outcomes, simulation modeling does not necessarily provide precise results that are completely reliable (eg, the exact number of infections or the precise course of an epidemic). Perfect prediction using simulation can rarely be achieved because it is impossible to build a model that fully replicates the real world, particularly when we describe a stochastic system as complex as infection transmission, which is influenced by human behavior, pathogen and host biological characteristics, and the health facility structure among many factors. Nonetheless, simulation modeling can help us to understand the relative effectiveness of different interventions, identify the risk of HAIs for different population groups, provide confidence intervals on the epidemic behaviors and, therefore, aid with decision making. IPC decision-makers using simulation models for decision-support must consider model assumptions and their relevance to the particular context in addition to carefully weighing the predicted benefits of interventions against the inconvenience, stigmatization, and costs they might engender.

A number of reviews have been conducted on mathematical modeling of HAIs in the 21st century. In 2006, Grundmann et al wrote the first literature review on HAI modeling, the study of which focused on explaining the capacity of models to enhance epidemiological understanding in hospitals. Their work was restricted to the detailed description of a number of publications. Nelson and colleagues recently carried out a similarly in-depth and limited in- breadth literature review on economic analysis applied to HAIs using dynamic transmission models. In contrast, van Kleef and colleagues published a systematic review on the overall trends in the application and development of mathematical models of HAIs over time. Last, Opatowski et al illustrated the overall progress of mathematical and simulation modeling of multidrug-resistant bacteria, spread in both the community and hospital settings.

Since these reviews were conducted, a significant number of simulation models, including agent-based models (ABMs) and hybrid models, exploring the dynamics of HAIs have been published. The application of simulation modeling of HAIs has grown rapidly, possibly owing to the recognition of this methodology's advantages and the increasing capabilities of computers. The current adoption and application of HAIs simulation modeling need to be consolidated and updated to facilitate the further development of appropriate models, enabling the investigation and evaluation of the best practice for IPC under different health care settings from clinical and economic perspectives. Therefore, we conducted a systematic review to establish (1) how simulation models have been utilized to investigate HAIs and their mitigation, (2) how these models have evolved over time, and to identify (3) gaps in their adoption and (4) useful directions for their future development.

Reviewed simulation modeling types

System dynamics (SD)–A top-down continuous simulation modeling method, which characterizes the structure of dynamic and complex systems, using stocks, flows, feedback, and delays within such systems to explore how the system structure determines the system behavior. Stocks (or “levels”) are defined as aggregation or accumulations of inflows and outflows over an interval of time. Flows (or “rates”) change a stock over time by adding to (inflows) or subtracting from that stock (outflows). SD models are well-known for their ability to depict non-linear relationships, which derive from the existence of feedback processes that exist in which actors within a system will later be affected by their actions. In this review, we also consider compartmental models from the mathematical epidemiology and ecology literature that describe the disease transmission dynamics and link them to aspects of health care facilities and provision of services that effect outcome. These models similarly take a top-down approach that often assumes continuous time, and they are implemented using differential equations.

Discrete event simulation (DES)–A process-based simulation method used for modeling the operation of a system as a discrete sequence of activities and events in time, characterizing and analyzing queuing processes and networks of queues, and solving problems of resource use. Events, entities, attributes, and resources are the key components in DES. Entities are passive individual objects that possess attributes. These attributes are unique characteristics or features, such as age and health status. Resources, such as those defined in DES, require time to provide a service to an entity, making other entities wait and form a queue. Entities consume resources while they experience events. However, the consumption of those resources does not depend on individual-level entity behavior. As entities use up resources they are indirectly competing with other entities in the queue. DES allows for capturing the effect of variability, stochasticity, and randomness of multiple elements within a system, but it does not explicitly model feedbacks or interactions between entities.

Agent-based model (ABM)–A bottom-up simulation method for modeling dynamic and adaptive systems with autonomous entities called agents and their environment. , The agents are described by their properties, actions, decision rules, and possibly goals, and interact with each another and the environment. They live in the environment, and they sense it. They decide what action to employ at a certain time on the basis of their own state, their own defined decision rules and the environment state (including other agents with which they interact). Agents can have explicit targets to minimize or maximize, and they can also learn and adapt based on experience. Agent-agent and agent-environment interactions result in the update of the internal state of the agents or a decision on their next actions. Similarly, the state of the environment can update. As agents and the environment interact and evolve or potentially co-evolve, micro- and macro-level patterns emerge. We also view similar microsimulation and individual-based models from the mathematical epidemiology and ecology literature as ABMs in this review, although in these models the entities are often only reflexive and do not make autonomous decisions.

Hybrid simulation models–A simulation modeling method that combines the methodological strengths of at least 2 different simulation modeling methods. We describe a number of designs for hybridizing simulation models based on the work by Morgan et al.

Sequential design–A design for combining 2 or more simulation modeling methods that can capture different parts/behaviors of the same system or at different levels of detail. The simulation models that are hybridized interact with one another in a way that information or data are passed from one model to the next model.

Enrichment design–A design for combining 2 or more simulation modeling methods to form a single model in which one method remains the core method that defines the system and other enhancing methods are transferred into and embedded within the primary method.

Integration design–A design for combining 2 or more simulation modeling methods to form a single model, which presents one coherent and concise view of the system, and captures interactive influences within the system.

Interaction design–A design for combining 2 or more simulation modeling methods in which individual models can operate independently but work together to capture interactive influences within the system.

Parallel design–A design for combining 2 or more simulation modeling methods that provide 2 potential representations of the same system, offering complementary insights of the system.

Table 1 provides an overview of the assumptions, inputs, outputs, and data dependency for each simulation modeling methodology. Other studies compare different aspects of these simulation methodologies more generally than in HAI but in greater detail, including Parunak et al, Phelan, Schieritz and Milling, Borshchev and Filippov, Rahmandad and Sterman, Siebers et al, and Scheidegger et al. , 27, 28, 29, 30, 31, 32

Table 1

Overview of the assumptions, inputs, outputs, and data dependency of SD, DES, and ABM

Feature	SD	DES	ABM
Assumptions	Entities within each stock are mixed homogeneously; simulation is deterministic.	Entities are passive and do not interact with one another or learn from or adapt to the environment, but they can be heterogeneous; simulation is stochastic.	Entities can be heterogeneous and autonomous decision-makers, who can learn and adapt to their environment; entities can interact with each other; simulation is typically stochastic.
Inputs	Stock and feedback and accumulation structures; initial levels of stock/sub-populations aggregated by particular characteristics; rates, which characterize the inflows and outflows of a stock.	Structure of queuing network; types of entities and resources (eg, HCWs, hospital beds and equipment), and their characteristics; time between entity arrivals, and number of entities per arrival; service time or delays.	Agent types and definitions in terms of their characteristics, possible actions and rules of behavior; initial number of agents; environment characteristics and rules; definition of agent-agent (eg, network), agent-self, and agent-environment interactions.
Outputs	Deterministic time series of population/stock levels and flows and insight into behavior of the system.	Stochastic time series of, and insight into, operational performance outputs such as queue lengths, utilization of resources, and frequency of events; tracking of individual entities.	Stochastic (typically) time-series of population and sub-population outputs such as number of entities in a specific state, frequency of actions, and frequency of events as well as state of the environment; insights into the system emergence behavior; tracking individual entities.
Data dependency	Objective data at aggregate levels supplemented by judgmental, subjective data, and informational links	Depending on simulation aims, these methods can be highly data-dependent because they model entities at the individual level and try to describe variations in their characteristics and other inputs.

ABM, agent-based model; DES, discreet event simulation; SD, system dynamics.

Methods

Information sources and search strategy

Pubmed, EMBASE, Cochrane library, ABI/INFORM Collection via ProQuest, Business Source Complete, and Scopus were searched from the date of inception through February 19, 2019. Results were restricted to peer-reviewed publications that were written in English. Search terms for HAIs were combined with search terms for simulation models as follows:

Infection OR infections

AND

Health care associated OR hospital acquired OR nosocomial OR HAI* OR HCAI*

AND

System dynamic* OR compartmental OR agent based OR microsimulation* OR discrete event* OR simulation*

All databases were searched identically. The detailed search strategy for each database is located in Appendix A. Reference lists of the previous literature reviews15, 16, 17, 18 were also searched for relevant citations.

Eligibility criteria

We included studies that had fulfilled all of the following criteria: 1) simulation modeling of the dynamics of HAI transmission, clinical and economic evaluation of preventions for HAIs, and/or the dynamics of antimicrobial resistance; 2) simulation models, including SD, DES, and/or ABM; 3) a primary focus on HAI transmission in health care settings, including hospitals, long-term care facilities (LTCFs) (eg, nursing homes and care homes), and/or medical centers.

Exclusion criteria

We excluded studies that did not involve either (1) human-to-human transmission or (2) human-environment-human transmission or that did involve (3) animal transmission of HAI or (4) pharmacokinetics and/or pharmacodynamics of antimicrobial drugs and/or molecular biological perspectives within the host (eg, molecular mechanisms of antibiotic resistance within the host, efficacy and/or side effects of antibiotics, and mode of action of drugs); or (5) within host immunity or strain competition only; or (6) community transmission of pathogens spread in the health care environment as well, in which the focus of the articles was community spread (eg, SARS epidemics); or (7) literature reviews that did not contain new primary studies. In addition, we did not include editorials or letters to editors.

Data collection process

Data were extracted for the included studies, categorized, and summarized in tabular format (Appendix Table A1).

Data items

We extracted key data to answer the objectives of this review. First, this contained the basic information of the studies (study title, authors, and year of publication). Second, because the main purpose of the review was to explore the existing use of simulation modeling for understanding HAI transmission and improving IPC in various health care settings from clinical and economical perspectives, we looked for the following codes: country of research, setting, type of simulation model, research theme, aim of the simulation model, pathogen, and inclusion of economic analysis. Additionally, because we were interested in how models of HAI transmission in health care settings were simulated to evaluate the effectiveness of IPC strategies, data on the type of intervention and the type of interactions (ie, patient-health care worker (HCW), HCW-HCW, patient-patient, patient-visitor, environment reservoir for transmission, interaction between health facilities and interaction between health facility and community) were also extracted. Additionally, to explore how different types of simulation models and hybrid models have been used, we looked into technical perspectives of these models, which included whether sensitivity analysis, simulation software, calibration, validation and verification, transferability, and generalizability were performed and how they were performed.

Results

Study selection

Figure 1 shows the process of identification, screening, and selection using the PRISMA flowchart. There were 606 records identified from electronic database searches and 25 records from other sources. After removing duplicates and reviewing the title and abstract of the remainder, full-text articles were retrieved for the retained 109 records to assess their eligibility. A total of 54 records were removed because they did not meet the inclusion criteria. An additional 13 studies were identified via reference screening of the existing systematic reviews.15, 16, 17, 18 Overall, 68 publications were included and reviewed in detail.10, 14, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 49, 50, 52, 53, 54, 55, 56, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 85, 86, 87, 88, 89, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115

Fig 1

PRISMA flow diagram.

Causative organisms modeled

Almost one-half of the included studies investigated the dynamics of methicillin-resistant Staphylococcus aureus (MRSA) in a health care setting (47%, 38 studies), followed by vancomycin-resistant Enterococci (VRE) and Clostridium difficile (CD), with significantly fewer studies (12%, 10 studies and 7%, 6 studies, respectively). Other pathogens have rarely been the subject of interest for studies in this field.

Country of research

One-quarter of the publications did not specify a particular country; however, of the studies that did specify a particular country, only 2 (3%) studied the health care setting in a middle-income country (South Africa and Thailand) and 3 (4%) reviewed an upper middle-income country (China). Most of the publications (68%, 46 studies) concentrated on HICs, of which nearly one-half were US publications (21 studies).

Types of simulation model

SD models accounted for 38% of the simulation models (26 studies). The first SD model of HAIs was developed in 1997, whereas ABM and DES models of HAIs were only introduced in 2005 and 2006, respectively (ie, nearly 10 years later). , Although ABM and DES models of HAIs were introduced nearly concurrently, ABMs were used much more frequently to model HAIs than DES, and they accounted for more than one-third of the reviewed models (38%, 26 studies).

Model hybridization

Hybrid model use has increased since they were introduced in 2007. Thirteen percent of the included studies (9 studies) adopted hybrid models, which combined 2 types of simulation modeling.37, 38, 39, 40, 41, 42, 43, 44, 45 Based on a toolkit of designs for hybridizing 2 types of simulation modeling proposed by Morgan and colleagues, we identified that 6 studies mixed ABM and SD models using either the enrichment, interaction, or integration , , approach. Two studies adopted a sequential design to combine SD and DES, , and one study used a SD model and a stochastic continuous time Markov chain model in a parallel design.

Sequential design

Van den Dool et al used a sequential design in which SD and DES was combined to capture different parts of the same system. This approach provided emergent insights as understanding of the system was enhanced. In their study, a SD model was first built to simulate an influenza pandemic in the community. DES was then adopted to simulate the transmission dynamics of nosocomial influenza in a LTCF. As the prevalence of influenza virus in the community influences the rate at which patients, health care workers (HCWs) and visitors introduce the virus when they enter the LTCF, the prevalence and the incidence of infections generated by the SD model were passed to the DES model. This hybridization improved understanding of transmission dynamics of nosocomial influenza in a LTCF by taking into consideration the impact of infection prevalence in the community on that LTCF. Wendelboe et al. reconstructed this hybrid model and validated it using the collected surveillance data for the period of 2006-2007 obtained from an active system of 76 LTCFs in New Mexico (the United States).

Enrichment design

In 2011, Barnes et al adopted an enrichment design to combine an ABM and a SD model to form a unified model. The study investigated how the interconnectivity and transfer of patients between various health care facilities influences the prevalence of HAIs at each facility. The SD model was simulated to determine the proportions of 3 patient states of infection (ie, susceptible patients, and persistently and transiently asymptomatic carriers), which formed a unique state for each health care facility. An individual facility was then modeled as an agent in a network of many health care facilities in the ABM. The role of the SD component was to generate the distinct characteristics of each agent, and it was embedded within the ABM method. Meanwhile, the emphasis was placed on the ABM component of hybrid model because it was responsible for addressing the objectives of the study as stated previously.

Integration design

In 2013, Sadsad et al designed a hybrid model by mixing SD and ABM into a single model in an integration design to look into MRSA transmission dynamics in a hospital at multiple-levels. The SD method was employed to simulate the flow of patient and HCW between different hospital wards and rooms represented as stocks. ABM was adopted to model transmission between patient agents mediated by HCW agents. During the modeling process, neither systems simulation methods were dominant, however, they were inseparable.

Caudill et al , also integrated a SD model and an ABM to form a single, unified hybrid model, which captured the intra-host dynamics of antibiotic-resistant bacteria and the inter-host dynamics of HAI spread among patients and HCWs. The ABM component facilitated the simulation of interactions between patients and HCWs. Individual patients and HCWs were represented as agents characterized by distinct properties and behaviors. The SD component simulated the changes in the internal pathogen population of each agent, called bacteria population vector over time, which form one of the elements driving the transmission probabilities during events of agent-interactions. This bacteria population vector of each agent defined the colonization or infection status of that agent. The status of the agents affected the transmission of bacteria between agents. Whenever interactions between 2 agents occurred or patient agents received the application or dosage of a specific antibiotic, the SD component was invoked to simulate the dynamics in the bacteria population vector within each agent and update its infection and/or colonization status. The ABM and SD component of the hybrid model were treated on an equal footing.

Interaction design

Kardas-Sloma et al used a previously developed ABM (ie, NosoSim) to simulate the spread of MRSA among patients and HCWs in a hypothetical intensive care unit (ICU). This model was coupled with a SD model, which simulated the transmission of MRSA in the community through hospital admissions and discharges. The hybrid model captured the interactive influences between hospital setting and community setting while the transmission within each setting was grounded in each method. The ABM and SD model adopted could operate independently or they could work together to enhance the understanding of the impact of overall reduction in antibiotic use upon MRSA selection in both settings.

Parallel design

Wang et al developed 2 separate HAI models using SD and a stochastic continuous time Markov chain, which offered 2 possible representations of the transmission dynamics of MRSA in a hospital. The hybrid model helped obtain complementary insights of the single system and revealed plausible explanations of the system's behaviors. This was achieved by the introduction of the SD model for the transmission of MRSA in a hospital followed by a stochastic epidemic model to check the important features, which had not been well illustrated in the other model. No interaction between the 2 models was observed.

The study of D'Agata et al applied the same approach to model the transmission dynamics of antibiotic resistant bacteria in a hospital setting from different levels of details. An ABM was used to model heterogeneous patient and HCW behaviors within a typical hospital setting and simulate infection spread. A corresponding SD model represented the system at an aggregate level that provided the interpretation for the behaviors of the ABM over a large number of simulations.

Sequential, enrichment, interaction, and integration designs of model hybridization have been useful for capturing different aspects/behaviors of the same system whereas a parallel design offers 2 possible presentations of the same system.

Sensitivity analysis

On completion of building a simulation model, it is important to evaluate how sensitive and/or robust the model is to various sets of initial conditions that we are using (ie, examination of the influence of varying parameter inputs on model results) because of the uncertainty of input parameter values and distributions for simulation model of HAIs. This process is called sensitivity analysis. Less than one-half of the studies included sensitivity analysis (47%, 32 studies). Of the studies that conducted a sensitivity analysis using one type of sensitivity analysis, univariate sensitivity analysis was the most common method (24%, 16 studies). Probabilistic sensitivity analysis (PSA) was the second most common method but to a significantly less extent (12%, 8 studies). PSA is generally regarded as a more rigorous method to explain uncertainty in the joint distribution of parameters, and is recommended in health economic evaluation guidance. In addition, 2 recent studies employed the combination of univariate and multivariate/probabilistic sensitivity analysis to investigate model sensitivity. , The number of studies conducting sensitivity analysis and the use of more sophisticated approaches have been increasing in recent years (Fig. 2 A).

Fig 2

(A) Use of different types of sensitivity analysis over time. (B) Inclusion of calibration, validation and verification process in simulation models of health care–associated infections.

Validation and/or confidence building and verification

The usefulness of a model and its results are particularly important to many stakeholders who use the results for decision-making or who are influenced by decision-making based on the models. Therefore, the ultimate goal of validation and/or confidence building and verification of a simulation model is to ensure the correspondence between reality and the implemented model to the degree, which satisfies the intended application or purpose of the model.

Extensive model validation had not been common practice in HAI simulation modeling; however, the percentage of publications including model validation increased until 2010 and has remained relatively constant (Fig. 2B). More than one-third of the included publications did not state any sort of validation (24 studies). Almost one-half of the simulation models that contained stochastic element (30 studies) were validated by using the single approach of internal validity, in which several simulations were performed to assess their stochastic variability. The lack of consistency in a model's results may cause the appropriateness of the investigated system or the strategy and/or policy to be questionable. Historical data validation, which uses a part of the collected data other than the data used for model building to test if the model behaves as the real system does, was found to be used as the single method of validation in 4 other studies (6%). Other validation methods were rarely used. Recent simulation models combined multiple validation methods to achieve a more thorough validation approach. , , , , , 52, 53, 54, 55, 56 With respect to the different simulation methods, approximately one-half of SD and hybrid simulation models depicted a validation approach, whereas a higher proportion was observed for the ABM and DES models (ie, 73% and 86%, respectively). More than one-half of the hybrid simulation models that included a validation method used combinations of validation approaches. This number occupied one-third of all of the models using mixed approaches to validation. Only 1 out of a total of 68 studies described how verification was conducted, using good documentation of the model building process, and randomly checking whether the simulated behaviors of selected agents of each type matched the intended behaviors of the conceptual model.

Model parameters and model calibration and/or model fitting

Parameters used for the simulation models came from published studies, assumptions, and/or real-world data obtained from clinical databases, observations, surveys, or were estimated directly from the data. Calibration has traditionally been considered as a method to adjust unavailable or unobserved parameters, such as infection transmission rates, to achieve a good fit with the data. Although the proportion of models that included some form of calibration is small (31%, 21 studies), this figure increased between 1997 and 2007 and has remained stable (Fig. 4). The models in this review used a number of calibration approaches: maximum likelihood estimation , ; the least- square criterion , ; Monte Carlo, Markov chain, and combinations of these methods. , , , , 62, 63, 64 In particular, McBryde et al used a combination of Bayesian estimation, the Markov chain, and the Monte Carlo for model fitting. Similarly, Sadsad and colleagues combined a scatter search algorithm and a least squares criterion for model calibration. Other studies compared model predictions with observed epidemiological data, , , , , whereas the remaining studies did not specify the model fitting approach.

Setting and interaction between settings

The simulation models of HAIs primarily depicted a single ward setting. , , Most of the models included in this review simulated transmission of HAIs in an ICU setting (25%, 17 studies) or a simplified hospital setting (32%, 22 studies) of which most lacked any further ward structure (12/22 studies, 55%). General wards were modeled in 5 studies (7%) of which 3 specified a particular type of ward (ie, an out-patient long-term hemodialysis, a dialysis unit, and a vascular unit). One recent study incorporated various types of hospital wards with a distinct nature into one model, including hospital wards designated as either general care, observation, and step-down, or the hospital wards were characterized as surgical and medical. Additionally, a small number of studies (5 studies, 7%) modeled the transmission dynamics of HAIs in LTCFs for pathogens, such as influenza, , , MRSA, and viral nosocomial gastroenteritis. Pediatric health facilities were considered in 3 studies. , ,

Most of the publications took neither the transfer patterns between health care facilities nor the transmission dynamics within health care settings into consideration although most ward or hospital-based simulation models included did not view the hospital as a closed system (eg, inclusion of hospital admission and discharge rates from and into community). Recently published studies incorporated the interaction between ICUs and general wards, or between general wards within a hospital. , , , , Donker et al was the first study to look at the impact of different referral patterns among various categories of hospitals on MRSA infection rates. One year later, 2 additional studies examined the interaction between the settings for MRSA. , Further, Lee and colleagues , explored the transmission of MRSA within a setting in which multiple hospitals, LTCFs, and the community interacted with one another.

Modes of interaction

Most of the studies asserted that the interaction between patients mediated via HCWs is the primary cause of HAI transmission in health care settings (75%, 51 studies). The remaining 17 studies (25%) did not specify any types of human-human or human-environment-human interaction that had been considered in their models. In contrast, a significantly smaller proportion of models in the review simulated other types of interactions within health care settings. In particular, both direct contact between patients or indirect contact via a contaminated environment were modeled in 24% and 22% of the included studies, respectively (16 and 15 studies), followed by direct contact between HCWs (16%, 11 studies) and visitor-patient contact (only 13%, 9 studies). Additionally, the inclusion of contact between the family caregiver and the HCW was a distinct characteristic of the models set up in the neonatal ICU. , This type of interaction is of importance and likely to happen within pediatric health care settings in which parents are often (if not always) involved in child care activities.

Software

Just over one-half of the studies in this review specified the software on which the simulation model had been built (53%, 36 studies). Table 1 shows that C++, MATLAB, AnyLogic, and NetLogo were the most popular. Although ABM was introduced much later than SD and DES, there was a growth in the number of the types of software available for ABM users. MATLAB, NetLogo, Anylogic, and Repast were used to hybridize different kinds of simulation models, mainly for mixing SD and ABM.

Types of health care workers modeled

Only one-quarter of the publications clarified the different kinds of HCW modeled. They mainly included doctors and nurses who are primarily responsible for the delivery of care in a health care facility, and therefore having the most frequent contact with patients (24%, 17 studies). Only a small proportion of the models simulated transmission caused by HCWs other than doctors and nurses (8%, 6 studies), which included peripatetic HCWs, , rogue HCWs, respiratory therapists, occupational therapists, speech therapists, physical therapists, admission personnel, auxiliary personnel and cleaning staff, and volunteers. Additionally, Jemenez and colleagues published a study in 2013 that created one of the most comprehensive social networks among patients and different types of HCWs in a simulated hospital, in which individuals had their own activity schedule.

Interventions for HAIs being modeled for effectiveness evaluation

The main theme of simulation modeling studies in HAIs has been to evaluate the effectiveness of various IPC strategies (87%, 59 studies). The intervention strategies being investigated in the studies included in this review were: hand hygiene (39%, 23 studies), patient isolation (27%, 15 studies), screening and antibiotic stewardship (22% for each type of intervention, 13 studies), decolonization (19%, 11 studies), and HCW cohorting (17%, 10 studies). Some studies assessed the effectiveness of integrating 2 different IPC strategies, including the effect of combining hand hygiene and decolonization for MRSA, isolation and screening for MRSA, and screening and contact isolation. A study published in 2015 used simulation modeling to conduct a more intensive assessment of the impact of mixing 4 different interventions. Similarly, another publication released 1 year later assessed the benefits of a “bundle” IPC strategy. Researchers have not extensively explored IPC measures, such as vaccination, patient cohorting, barrier precaution, environmental disinfection, and referral patterns.

Economic evaluation

A minority of the included publications included an economic evaluation of HAIs (10 studies, 15%). In 2009, a model first adopted DES to conduct a cost-effective analysis based on actual data from 2 hospitals in the United States. This study strongly suggested the association between length of stay and HAIs, which had been ignored in previous publications. , Recently published studies paid more attention to the economic aspect of HAIs. They have estimated cost-effectiveness for different IPC strategies and investments, mainly for MRSA , 71, 72, 73 , , followed by CD. , Economic analyses were carried out for a single intervention (ie, hand hygiene, , isolation, vaccination, , patient room design), combination of 2 , or 3 interventions, , and a bundled strategy. It can be clearly seen that most studies focused on the cost-effectiveness evaluation of hand hygiene, screening and isolation. Table 2 shows a summarized description of the included studies’ economic analysis for HAIs.

Table 2

Description of the studies that included economic analysis

First author	Year of publication	Pathogens	Model types	Setting	Type of economic analysis	Interventions
Hagtvedt59	2009	MRSA, VRE	DES	ICU	Cost-effective analysis	Hand hygiene, isolation and combination of measures
Hubben72	2011	MRSA	DES	Entire hospital	Cost-effective analysis	Selected vs universal screening
Greer70	2011	Pertussis	ABM	NICU	Cost-effective analysis	No vaccination vs vaccination
Robotham85	2011	MRSA	ABM	ICU	Cost-effective analysis	Screening, isolation, decolonization and combination of measures
Gurieva73	2013	MRSA	DES	ICUs and general wards	Cost-effective analysis	Screening, isolation and combination of measures
Nelson82	2016	Clostridium difficile	ABM	Entire hospital	Cost-effective analysis	Bundled measure, including testing, isolation, hand hygiene, contact precautions, soap and water for hand hygiene, and environmental cleaning
Robotham86	2016	MRSA	ABM	Entire hospital	Cost-effective analysis	Options for MRSA screening for admitted patients (no screening, checklist-activated screening, and high-risk specialty-based screening), isolation, decolonization, and combination of measures
Shin56	2017	MERS	SD	Entire hospital	Cost-effective analysis	Patient room design
Stephenson87	2017	C difficile	SD	Entire hospital	Cost-effective analysis	Vaccination strategies
Luangasanatip71	2018	MRSA	SD	ICUs	Cost-utility analysis	Hand hygiene

ABM, agent-based model; DES, discreet event simulation; ICUs, intensive care units; MRSA, methicillin-resistant Staphylococcus aureus; MERS, Middle East respiratory syndrome; NICU, neonatal intensive care unit; SD, system dynamics; VRE, vancomycin-resistant Enterococci.

Transferability and generalizability

Because of economical, logistical and theoretical benefits, it is important for model users to understand how to enhance model transferability and generalizability during model development. However, as models imperfectly represent real systems and are contextually constrained during their development, care needs to be taken when transferring and generalizing an existing model to avoid unintentional misapplication. Most of studies included in this review did not discuss the transferability and generalizability of the developed simulation models (78%, 53 studies). Of the studies that state these aspects, they briefly discussed the possibility of transferring their simulation models to assess HAI transmission dynamics for different pathogens, , , in different health care settings, , , and to evaluate the effectiveness of different sets of interventions. , However, a methodology for model transferring or generalizing, rather than modification of parameter values, model setup and assumptions, was not clearly explained.

Benefits of using simulation modeling

Only 5 studies state the benefits of using simulation modeling in health care (7%). The reasoning outlined in the studies to rationalize the use of this method included time, cost, and practical and ethical considerations of experimental or observational research methods, such as randomized controlled trials. , Another reason was the complexity of transmission dynamics, spread and resistance of HAIs, which involve numerous interdependent and dynamic interactions and cannot be completely captured by epidemiological studies. , , The advantages of ABM over other simulation modeling methods were also discussed in 4 articles, mainly emphasizing its capability to simulate the heterogeneity of patients and behaviors of HCWs in health care settings and their contact networks. , , , These studies indicated that ABM was the most appropriate for modeling an ICU in which the population size is small and patient turnover is high. Neither a clearer explanation of the pros and cons of each simulation modeling nor when to combine them and what the benefits of doing so were found in the reviewed studies.

Discussion

How have simulation models been used to enhance the understanding of HAIs and IPC?

MRSA was the predominant pathogen modeled, followed by VRE and CD to a significantly lesser degree. As MRSA accounts for high rates of morbidity and mortality, and can lead to metastatic or complicated infections such as sepsis or infective endocarditis, it remains a global health issue. Similarly, VRE has been a significant cause of HAIs, likely affecting the most vulnerable patient groups and accounting for significant mortality rates with prolonged LoS and therefore increased health care costs. Both of these pathogens have become the subjects of national IPC policies and the targets of national surveillance systems in a variety of HICs.91, 92, 93 Therefore, it is understandable why MRSA and VRE have been the pathogens of interest in many simulation models for HAIs.

The problems of HAIs in LMICs in which the burden is significantly higher than HICs are rarely addressed in the literature and particularly in simulation modeling studies. The prevalence of HAIs in LMICs is at least double the prevalence in Europe. Additionally, the incidence of HAIs acquired in ICUs in LMICs triples the incidence in the United States. However, our review found that only a minority of simulation models for HAIs in LMICs were developed.

ICUs have remained the subject setting of several simulation models because they are one of the most dynamic and complex areas in a hospital. Simulation models for HAIs have also become more complex in terms of the settings being modeled. Earlier studies generally modeled a single ward (usually an ICU) or a simplified hospital lacking of any further ward structure whereas more recent studies were likely to incorporate different types of wards (ICUs and general wards), as well as consider the transmission across health facilities (eg, mainly between hospitals), and the community. Future studies could investigate interactions with LTCFs, other types of health care facilities, and the community to provide a more realistic estimate of HAI incidence and prevalence, and the effectiveness of IPC policies. Pediatric settings were rarely considered although pediatric patients have higher rates of viral lower respiratory tract infections and bloodstream infections than adults, especially those younger than 2 years of age and those demanding care in neonatal ICUs and pediatric ICUs.

As the most popular transmission routes of infections in health care settings are via the transiently colonized hands of HCWs and/or contaminated medical equipment and the environment, modeling interactions between patient and HCW has dominated this field of research, followed by the environmental reservoir for transmission although to a much lesser extent. Simulation modeling studies have hardly considered direct HCW-to-HCW contact or interactions between visitors/caregivers and patients. Visitors and/or family caregivers can play a very important role in infection transmission in a health facility, especially in settings such as pediatric or geriatric health facilities where patients often need extra care. In many cultures including Asian countries and LMICs, having visitors and caregivers on a regular basis is common practice and sometimes encouraged owing to a considerable shortage of staff and a need to reduce medical costs to patients. Because visitors and caregivers are also more mobile than patients, they are both highly susceptible to contracting infections and potentially able to transmit pathogens to various locations inside and outside as the hospital.

When to use which simulation modeling methodology

The application of 3 types of simulation models to investigate HAIs has greatly changed over time. SD is suitable for investigating the long-term behavior of the system containing large patient populations, which are considered to be homogenous and therefore aggregated into compartments. Thus, it is useful for macro-level modeling to reflect long-term consequences and discover long-term solutions that may provide effective aids in policy decision making at a high level. Although SD has long been used to analyze HAI dissemination in hospitals and IPC policies, it could not address the spatial detail and microstructure of a health care facility, the complexity and heterogeneity of contact networks within a health care setting and the stochasticity of interactions within such networks.

By contrast, ABM has been found to be significantly helpful in overcoming the limitations of SD, which may explain the increasing use to model HAIs in recent years. It is easier and thus preferable for modeling the heterogeneity of a small population like an ICU rather than a large population setting. Health care settings in general and ICUs, in particular, are spatially intricate environments where complex interactions between specific sets of individuals are a key driver of transmission. Not every primary physician, consultant, and nurse see every patient, leading to a highly heterogeneous social and contact network. Diagnostic uncertainty (ie, whether an individual is infected is not always known) also complicates the transmission of HAIs. This accentuates the importance of impacts of stochastic interactions and chance events upon the transmission and spread of HAIs. ABM can also help understand the influence of different patient referring and transferring patterns among health care facilities within a network owing to variations in their geographical locations, policies, services provided and variations in individuals’ decision. A limitation of ABM is the requirement of reliable and detailed data for model building and validation, which are not always readily available. Higher levels of behavioral detail produced by ABM causes greater computational intensity, and difficulty in performing model parameterization and extensive uncertainty analyses, which are essential for reliable predictions. As ABM and hybrid models become increasingly popular, the adoption of more sophisticated methods and mixed methods for sensitivity analysis, calibration and validation were more frequently observed in more recently published studies.

Similar to ABM, DES allows incorporation of detailed patient attributes and is well-suited for modeling the procedure of activities that patients need to progress through. However, unlike ABM, DES does not consider social contacts and interaction among individuals, and therefore, transmission of infections needs to be simulated indirectly in a DES model. DES cannot model individual-level behaviors such as learning, adapting and autonomous decision-making as ABM does. Nor can it capture feedbacks in a system as SD does. Therefore, it is less satisfactory for simulating transmission of pathogens, possibly accounting for the less frequent application of this simulation modeling method in HAIs in comparison with the use of SD and ABM.

The adoption of hybrid simulation models has become increasingly common. Because all 3 simulation modeling methods have different benefits, limitations, strengths, and weaknesses, mixing methods potentially overcome some of the drawbacks faced by using a single approach and/or provide more plausible explanations of a problem, which a single method on its own could not handle. For example, SD is useful in providing a holistic view of the feedback dynamics of HAI transmission in a complex health care system but cannot take account of the heterogeneity of individual patients and HCWs, and the stochasticity resulting from their behaviors and interactions, which are the distinct features of ABM. As health care systems are highly complex, dynamic and interconnected, HAIs and other problems in the context of health care gained from different simulation modeling methods may benefit from the complementary view gained from using multiple simulation modeling methods together. However, a clear framework and philosophical foundation for hybridization have not yet been established in any of the reviewed publications.

Few studies included in this review explicitly explain why they choose one methodology over the others to answer their research questions. Therefore, the rationale underlying the use of different simulation methodologies in HAIs is still not clear. The choice of simulation methodology should be problem-driven and depend on the research objectives and the availability of data. Future modeling studies should be encouraged to include explicit explanation for the selection of a specific simulation methodology. This would provide insights for researchers and modelers in this field with respect to the different uses for each simulation methodology. Further, a full framework for choosing a simulation methodology should be broached in future research.

Implication

This review provides an overview of the development and application of systems simulation modeling in HAIs from which gaps of research in this field can be identified. First, the transmission patterns of HAIs in LMICs require further studies because they are likely to be dramatically different from those in HICs owing to many factors such as poor infrastructure, insufficient environmental hygiene conditions, different staff cohorting, shortage of HCWs, the knowledge and compliance of HCWs to IPC measures, overcrowded health care facilities, absence of comprehensive IPC guidelines and policies, lack of procedure, and different antibiotic prescribing and referral patterns. Second, pediatric hospitals and other types of health care setting-like LTCFs, as well as interactions between settings were not extensively investigated. Furthermore, understanding of patient sharing and referring networks among health care facilities driven by operational and financial alliances needs to be improved. Third, the number of studies adopting hybrid simulation models are still limited, possibly because of the unavailability of clear guidelines and frameworks for hybrid model development. Because it is argued that most, if not all, real-world problems tackled with simulation modeling cannot be solved by SD, DES or ABM alone but require a combination of 2 or all of them, a hybrid model resulting from this combination expectedly offers different perspectives of a problem and generate more insights, which will provide better understanding and greater support for decision-making. The use of simulation modeling for economic analysis of different IPC measures and strategies has increased but is still relatively scarce. The application of this methodology to evaluate the cost-effectiveness of various IPC strategies is promising in a sense that it can appropriately guide and prioritize the allocation of limited resources and funds. Additionally, an understanding of other kinds of interactions in the health care setting apart from interactions between doctors and/or nurses and patients is insufficient. Last, the evaluation of clinical and cost effectiveness was only conducted for a number of commonly used interventions like hand hygiene, isolation, and screening, further investigation on other IPC measures and a combination of different strategies is imperative to determine best practice in various health care settings. Models can also be developed to simulate coordination and collaboration among health facilities to assess the impact of a regional IPC program.

Conclusions

The review aims to consolidate and update the development and application of systems simulation modeling in studying HAIs. It can help guide further development of simulation models, especially hybrid models, to target gaps in knowledge in this field of research. The results of this review indicate that the complexity of simulation models for HAIs, in terms of the level of details of health care settings and interactions being modeled and methodological designs, significantly increased over time; however, the context predominately remained focused on the transmission dynamics of MRSA in hospitals in HICs, rather than in other types of health care settings, such as LTCFs or in LMICs. Additionally, the overview of existing simulation models in HAIs can facilitate and direct researchers to useful areas for further research such as transmission of HAIs in health care settings other than hospitals and across different types of settings. Further development and application of hybrid simulation models could help to secure further insights into HAIs.