The elephant in the 2020–2021 hospital room has been COVID-19. Over the last year, SARS-CoV-2, the virus responsible for COVID-19 rose from relative obscurity to over 120 million detected cases worldwide (as of 16 March 2021) [6,7]. The meteoric rise in cases has been paralleled by a similar rise in modeling research. The early focus of modeling centered around community transmission, healthcare utilization needs, and projections of interventions designed to maintain health system integrity [8–12]. In addition, forecast and simulation studies sought to explore the risk of nosocomial transmission of SARS-CoV-2 among healthcare personnel and patients [13].

As the pandemic evolved, significant attention has been paid to modeling nosocomial transmission of SARS-CoV-2 in long-term care facilities (LTCFs). Early in the pandemic, LTCFs proved to be particularly high-risk settings due to high concentrations of individuals with high-risk comorbidities and the congregate living arrangement. Consequently, a number of modeling studies sought to understand LTCF bed needs, testing requirements for preventing transmission, and optimal vaccine deployment strategies [14■,15–18].

The CDC estimated that in 2019 there were over 320 000 cases of methicillin-resistant Staphylococcus aureus (MRSA) in the United States resulting in over 10 000 deaths [3■]. MRSA colonization is associated with susceptibility to invasive infection, highlighting the benefit of identifying who might benefit from decolonization regiments. One study aimed to make screening more cost-effective by developing a clinical prediction rule for identifying patients at high risk of being colonized with MRSA [19■]. The model showed that a negative predictive value of 99.4% can be maintained when just 25.4% of patients are screened. In another study, the possibility of MRSA superspreading events within and between healthcare facilities was shown in an analysis of an extensive contact network of patients, emphasizing the impact and importance of screening on admission [20■■].

Models also addressed the effectiveness of interventions for preventing MRSA transmission. In a large-scale study of 108 Veterans Affairs acute care hospitals, Khader et al. [21■] found contact precautions for MRSA patients reduced transmission by 47%. Furthermore, genomic surveillance of MRSA is likely to be cost-effective based on the reduction of MRSA acquisitions in the year following an admission [22■].

MDROs are a frequent contributor to HAI [1]. The CDC estimated that for 2019 there were over 197 000 extended spectrum beta-lactamase Enterobacteriaceae infections resulting in over 9000 deaths [3■]. Modeling of epidemiological data for multidrug-resistant Klebsiella in Europe suggested that nosocomial transmission and hospital-based antibiotic consumption were the major drivers of transmission at a national level [23■■]. Thus improvements in infection prevention and antimicrobial stewardship may impact the emergence and spread of MDRO in both the hospital and community.

An important avenue of investigation focuses on how antibiotic resistance emerges as a consequence of clinical care. It has been suggested that the evolutionary rate for developing antimicrobial resistance is growing more quickly due to a nonlinear coupling between increased antibiotic use and the emergence of antibiotic resistant strains. However, it is unclear whether the observation of increased rapidity of antimicrobial resistance is simply a consequence of increased surveillance [24]. Models have also explored the potential impact of bystander exposure in which nonpathogenic bacteria in a patient develop drug resistance after they are exposed to antibiotics that are targetted for an invasive pathogen. In one analysis, up to 48% of bystander antibiotic exposures might be avoidable in the outpatient setting [25]. However, more studies are needed to quantify the relationship between bystander exposure and clinically meaningful antibiotic resistance [26].

Clostridioides difficile infection (CDI) is a frequent complication of medical treatment. In 2017, nearly 224 000 people in the United States required hospital care for CDI and at least 12 000 people died [3■]. Models of CDI have been used to assess the amount of transmission occurring within the hospital system, and thus offer perspective on the effectiveness of prevention and control efforts. For example, a combination of electronic health record and sequencing data lead to an estimate that 21% of CDI cases were acquired from another patient due to overlapping stays or via spores left behind on discharge [27■■]. This study also demonstrated strain dependence in both the transmissibility and durability of spores.

It has long been recognized that there is significant heterogeneity in transmission of CDI. This may be due to host, pathogen, or environmental factors. Models can be used to tease apart the causes of heterogeneity. For example, a recent modeling study using C. difficile surveillance data found that different wards in the same hospital can exhibit different nosocomial transmission dynamics [28].

A challenge in healthcare epidemiology is quantifying the benefit of specific control interventions such as the implementation of contact precautions. As more detailed data are made increasingly available, studies are beginning to place bounds on the impact of contact precautions and how it varies by organisms [21■,29].

An intriguing complexity of infectious disease dynamics is the inter-relationship between transmission pathways for multiple pathogens. For example, a model of the MRSA surveillance and prevention strategy in the Veterans Affairs system quantified how the impact of interventions designed to reduce the spread of MRSA may have also prevented a substantial number of infections from other organisms [30].

Building upon success in the control of nosocomial influenza transmission through targeted vaccination campaigns, there has been increased interest in vaccination as a tool for combatting other HAIs [31,32]. In advance of a promising CDI vaccine, Toth et al. quantified the indirect effects of a potential vaccine. This study suggested that between three and 16 CDI cases per 1000 vaccinated patients could be averted, including one to five cases among unvaccinated individuals [31].

Another approach to HAI prevention is decolonization of organisms that can precipitate invasive infection. The cost-effectiveness of new antibiotics, antiseptics, and other technologies that decolonize carriers of drug-resistant organisms appears markedly enhanced when the impact on community transmission is incorporated [33]. This poses an intriguing preventive strategy but would require reassessment of approval mechanisms for novel decolonization technologies.

Studies continue to show the benefit of antimicrobial stewardship programs in reducing HAI. For example, by curtailing nonspecific use of antibiotics, cases of CDI have been lowered by over 30% in some hospitals [34]. Several studies characterized the importance of coordinating antimicrobial stewardship programs among inpatient, outpatient and LTCF for reducing CDI and MDRO burden [35–37]. Other studies show the potential value of prediction algorithms for identifying patients on empiric antibiotics who have a very low risk of having a bloodstream infection [38].

The development of antimicrobial resistance is enhanced by antibiotic selection pressure; thus, there is increased interest in elucidating the relationship between the clinical use of antibiotics and emergence of antimicrobial resistant strains [39]. A recent model by Hansen et al. [40] suggested a minimalist strategy to exploit competitive suppression of selective strains over resistant strains so that chronic control of both invasive infection and resistance can be achieved.

The widespread availability of genomic data coupled with the increase in computational power has resulted in rapid development of methods that have significantly improved our ability to identify transmission links among patients. For example, a study by Eyre et al. [27■] developed a novel approach for integrating pathogen genomic data into an agent-based model for CDI transmission in a hospital setting. This methodology permitted inference of patient-to-patient transmission links while accounting for genetic recombination, readmission and additional routes of transmission such as persistent environmental contamination.

The COVID-19 pandemic has led to additional methodologic improvements for incorporating genomic data into epidemiological models. A study by Meredith et al. [41] used genomic data and epidemiological analyses of patients in a hospital to identify clusters of SARS-CoV-2 transmission. The results helped clinical, infection control, and hospital management teams to improve interventions and patient safety. Similarly, a study by Rockett et al. [42] coupled genomic analysis with an agent-based model to identify the ratio of local versus imported acquisition of SARS-CoV-2 in Australia.

Several studies have shown the use of machine learning to identify patients at high risk of acquiring various types of HAI. For example, one study developed a model to predict the probability of a urinary tract infection caused by a MDRO [43]. Another study predicted patients at high risk for central line-associated bloodstream infections (CLABSI) [44]. The risk of surgical site infections has also been explored [45]. The diversity in recent studies illustrates the broad applicability of machine learning methodologies.

The COVID-19 pandemic has highlighted disparities in outcomes across socioeconomic status, residential neighborhood, race, and ethnicity [46,47]. Models can help characterize structural factors that lead to inequity. Epidemiologic and population-level modeling studies have explored the extent to which inequities arise from variability in hospital care, heterogenous exposure in the workplace, higher rates of comorbidities in certain subpopulations, and geographic localization of intergenerational households [48–50]. While studies have begun to explore disparities, additional modeling studies are needed to translate these descriptive findings into actionable interventions such as improving access to healthcare or addressing mistrust that some communities have of the medical profession [51].

Predictive models for HAI should continue to explore the specific impact of demographic factors such as sex and race. Although a number of studies include demographic variables in the model building process, causal dependencies can be challenging to illicit. For example, a study identified black race as a predictor for CLABSI risk, but it was unclear whether this relationship was due to self-identified race serving as a proxy for socioeconomic factors [44]. In addition, the effects of implicit bias on prediction models deserve more attention. In particular, implicit biases may affect the selection of training data and thus bias predictions.

Despite this era of seemingly endless computational resources, challenges of model generalizability, and applicability abound [52,53]. A common problem identified in models focused on risk estimation is that they are invariably trained for a specific time and location. However, inevitable changes in epidemiology, treatments, or patient demographics can necessitate retraining or rewriting a model.

Internal validation of models requires a balance between accurate parameterization and overfitting. Thus, there are trade-offs between a time-invariant model that includes minimal site-specific information and a complex microsimulation model for which data might be lacking to support the specific assumptions. One approach for addressing this challenge maybe to use machine learning algorithms to parameterize mechanistic models tailored to our best scientific understanding of the spread of disease [5].

Beyond internal validation, barriers to implementation of algorithms include assurance of quality standards that are transferable across locations. For example, resistance patterns to antimicrobials in urinary tract infections can generally differ from region-to-region, and even from ward-to-ward within a hospital [43]. Encouragingly, many studies use separate hospital cohorts for external validation. However, careful consideration of the external validation is needed before broad application to a new health system. For example, training and external validation of a model using data in a single country may not translate well to other countries. Similarly, results based on data from urban areas may not apply to more rural areas. Additional studies are needed to evaluate how technologies such as propensity score matching can help evaluate the generalizability and safe clinical application of models beyond the time periods and populations in which internal and external validation was conducted [54].

Finally, a critical gap exists between statistical validation of models and the implementation of these advances for clinical or public health decision-making. All too often well developed models are not deployed to address the problems they target; rather they languish in academic journals collecting dust. To transition models from a theoretical concept to practical implementation requires communication among modelers, healthcare workers, healthcare administrators, and public health officials [55,56]. In particular, the rapid, high stakes collaborative effort that brought modelers, hospital administrators, and public health officials together to manage the COVID-19 pandemic has shown that greater emphasis on the deployment process is needed to efficiently and effectively address future outbreaks [57,58■].