Near–real-time evaluation of infection prevention response effectiveness

A major challenge in the practice of infection control is that when an increase in potentially preventable infections is identified, a bundle of interventions is developed to address the problem. Over time, the focus of the intervention improves, whether it be incidence of central-line–associated bloodstream infections or system-wide Clostridioides difficile rates. Because of the rarity of events and the natural tendency for regression to the mean, it can be difficult to determine whether the interventions deployed were effective, specifically which bundle elements were effective and should be retained and which were ineffective and therefore an inefficient use of resources. The learning healthcare system framework provides a path for continuous re-evaluation of infection prevention interventions and for generation of real-time data to inform the impacts of implementing (or deimplementing) different interventions. Thus, real-time electronic data tools represent major opportunities for advancing the science of clinical infection prevention practice, for improving care, and for promoting evidence-based deimplementation of ineffective interventions.

Whole-genome sequencing

Traditional surveillance methods of detecting suspected outbreaks relies on pathogen similarity. Often, isolate antibiograms are compared when an outbreak or transmission is suspected. However, this approach has limited sensitivity and specificity. Sensitivity can be limited due to genetically related pathogens having different antibiograms, and specificity can be limited due to distinct pathogens with similar antibiograms. This discordance can create noise and misclassification when using traditional surveillance tools and traditional epidemiologic associations to detect and evaluate outbreaks.

Genomic methods offer the ability to discriminate whether a common source exists between isolates and therefore whether transmission of a pathogen has likely occurred. ^{Reference Sundermann, Chen and Kumar7} Once reserved for only investigations of major outbreaks, genomic tools are becoming increasingly more common in infection prevention practice. Whole-genome sequencing (WGS) has replaced pulsed-field gel electrophoresis as the gold standard in genomic typing. Using WGS, investigators can show genetic relatedness of pathogens by comparing individual mutations in their genomes.

When is the best time to deploy WGS? Given historically high costs and limited availability, the traditional use case in infection prevention has been in response to suspected transmission events or to confirm or refute the presence of an outbreak after it has already occurred. Often, an outbreak investigation is initiated by infection prevention after a cluster of cases is identified or after questions are raised by a clinician. An infection preventionist (IP) may create line lists of suspected patients involved, perform audits, generate hypothesis, and initiate an intervention. Only then would WGS be applied to identify which isolates are part of an outbreak. However, such an approach is error prone, often misidentifying suspected transmissions and outbreaks ^{Reference Pfaff, Richmond and Sabin8} and failing to identify true outbreaks. ^{Reference Sundermann, Chen and Kumar7}

Decreasing costs of WGS and wider availability of the technology through public health, academic, and commercial laboratories has enabled some institutions to create WGS surveillance programs and/or to perform sequencing of isolates as they are cultured to detect or rule-out transmission events. These systematic approaches to applying the technology overcome the limitations of reactive sequencing and move away from relying on clinicians and epidemiologic associations to detect outbreaks. Although very promising, systematic applications of WGS have not yet been widely adopted in real-time settings and most of the data about their effectiveness comes from retrospective research investigations. ^{Reference Sundermann, Chen and Miller9}

Genomics and automated infection detection

WGS can augment automated infection surveillance tools that aim to detect outbreaks and transmission events. Tools such as WHONET-SaTScan use geotemporal aspects and antibiograms to automatically detect potential transmission events. Using the EHR, such tools identify pathogens with similar antibiotic sensitivities that were collected from similar locations and within a relevant timeframe. They can be used to estimate a probability of a transmission event or outbreak. Although it is intriguing to improve upon less systematic approaches to identify transmission events and clusters, the application of the geotemporal approach in a healthcare setting has a high false-positive rate that must be considered before it is implemented. ^{Reference Stachel, Pinto and Stelling10} For example, one study applying the WHONET-SaTScan detected 168 “outbreaks,” of which only 6 (3.5%) were considered to be true transmission events. The geotemporal clustering approach can also fail to identify transmission events, particularly those that occur during procedures (eg, from shared contaminated devices) that are performed in separate locations or on separate days and are therefore not detected. Augmenting geotemporal clustering with WGS surveillance could improve upon epidemiologic linkage alone to improve the positive predictive value of outbreak detection. Once identified, EHR algorithms could then be applied to determine potential routes of transmission to facilitate outbreak investigation and response.

In addition to its utility as a standalone tool for improving evaluations of potential transmission events, a WGS surveillance program coupled with automated infection surveillance tools could significantly assist infection prevention efforts in outbreak investigations. Under current typical circumstances, an infection prevention investigation requires intensive chart review of potentially involved patients. Under a hybrid approach that leverages clinical and genomics data, automated algorithms could extract rich EHR data and instantly analyze it to identify epidemiologic linkages and potential routes of transmission. Previous studies deployed this hybrid approach using a case–control design applied by a machine learning (ML) algorithm to identify associations that warrant additional evaluation. In these studies, cases were defined as all genetically related patient isolates detected by WGS surveillance and controls were all other patients in the hospital. ^{Reference Sundermann, Chen and Kumar11,Reference Sundermann, Chen and Miller12} The ML algorithm was able to successfully identify plausible routes of transmission for many of the outbreaks, which limited the need for an intensive, manual chart review process and allowed for a more streamlined outbreak investigation. An additional benefit was the ability of the ML algorithm to detect potential routes of transmission that might otherwise be missed by a fully manual case review process.

Technological barriers and realities of machine-learning algorithms

Although retrospective studies are intriguing, these advanced data science strategies have not been widely adopted for real-time, operational use. Theoretically, a major benefit of such informatics strategies is the lack of reliance on human resources. However, the reality is that such ML algorithms require labeled training data before they can be applied. Creation of training data sets requires substantial input, either based on data from earlier outbreaks that have been manually adjudicated or from infection prevention experts providing manually reviewed cases from a current investigation. In addition, ML algorithms must also undergo a validation step with a data set that is separate from the training set. The validation data set allows investigators to test the reliability and accuracy of the algorithm in a setting not biased by its training data.

For example, a ML algorithm to detect transmission events in possible outbreak investigations was trained using retrospective data from well-defined outbreaks with known routes of transmission. ^{Reference Sundermann, Miller and Marsh13,Reference Miller, Chen and Sundermann14} In the example of the outbreak investigation algorithm, the investigators applied their tool on subsequent WGS surveillance data and validated its findings on how the outbreaks were transmitted. ^{Reference Sundermann, Chen and Kumar11} In the future, once algorithms are trained and validated with historical data, some of these current challenges may be mitigated; however, ongoing maintenance and revalidation will be required to support these innovations.

Advancing the science of infection prevention by improving identification of clinical trials eligibility and facilitating outcomes assessments

Studies in other fields suggest that automated infection surveillance systems can be used to improve identification of patients eligible for clinical trials and to improve enrollment. ^{Reference Penberthy, Brown, Puma and Dahman15,Reference Ni, Bermudez, Kennebeck, Liddy-Hicks and Dexheimer16} Although this approach has not been used in infection control, early identification of patients with HAIs through automated infection surveillance processes provides an opportunity to develop and rapidly test different interventions and to assess their efficacy. After enrollment in a clinical trial, automated surveillance tools can be used as an objective measure of outcome assessment, potentially increasing clinical trial access by reducing the need for specific research expertise at participating centers. Fully automated surveillance of surgical-site infections to compare the effectiveness of oral antibiotics versus standard of care on outcomes following colorectal surgery using National Surgical Quality Improvement data is already underway. ^{Reference Apte, Moloo and Jeong17} Thus, advancing electronic detection strategies may not only improve case detection and reproducibility but may also be leveraged to advance the state of the science underpinning infection control practice.

Early warning systems and detection of rare and late healthcare-associated infection transmission events

Current HAI detection systems are inherently focused on near-term outcomes, and although some HAIs may cause limited harm early, substantial and potentially preventable harm may occur many years in the future. Although not currently possible, an artificial intelligence system with broad access to longitudinal EHRs across systems (and ideally genomics data) might be able to identify infections with a longer asymptomatic period early in the disease course or the transmission of novel infectious syndromes. For example, SARS-CoV-2 was not identified until substantial community spread had already been occurring for some time. ^{Reference Jorden and Rudman3} Early identification and pre-emptive treatment could avert adverse impacts and could facilitate rapid deployment of infection prevention resources to areas of greatest need.

Within genomics, an example of a centralized data system is PulseNet, which was created by the CDC and combines greater than 80 governmental laboratories across the United States. ¹⁸ Each individual laboratory sequences high-impact, potentially foodborne-related pathogens and submits the data to the CDC. This approach enables the CDC to detect multistate outbreaks that would have otherwise been undetected by individual hospitals or laboratories, which facilitates the public health response.

As EHRs are increasingly linked across healthcare systems, standardized and automated surveillance tools can also be used as part of an early warning system to detect seemingly unrelated clusters earlier, to allow for planning and implementation of interventions to avert additional infections, and to improve upon case reporting that is detected by frontline clinicians alone. For example, prior to implementing an automated infection surveillance strategy for adenovirus-induced hepatitis in children, the New York City Department of Health and Mental Hygiene rarely received reports of the illness. In April 2022, they initiated an automated infection surveillance system of emergency department visits that searched for certain International Classification of Disease, Tenth Revision (ICD-10) administrative codes and key words. Through this simple alert system, the health department were informed of ∼5 potential cases per week. These reports were then reviewed by disease investigators to determine whether the patient met the criteria for a patient under investigation (PUI). Within a few months, at least 10 PUIs had been identified in New York City (personal communication).

Advanced detection technologies may be especially useful for identifying extremely rare but “never miss” events such as HAIs that typically manifest many years after exposure (eg, HCV transmissions during endoscopies) ¹⁹ or HAIs that occur outside the healthcare system where the exposure occurred (eg, CRE transmissions from outpatient endoscopies that manifest as sepsis). Each of these scenarios has its own identification challenges; however, all are characterized by difficulties linking an event to the relevant exposure for a variety of reasons, whether that be time or follow-up care in another setting.

For example, viral infections, such as hepatitis C virus (HCV) are rarely transmitted through endoscopy procedures. During a large cluster of HCV infections in New York City, ^{Reference Branch-Elliman, Weiss, Balter, Bornschlegel and Phillips20} an outbreak was identified and reported due to a confluence of events: a high rate of acute HCV infections leading to inpatient hospitalizations at 3 different hospitals on the same weekend, simultaneous with the endoscopist who performed the procedures being on call at all of them. If the endoscopist performing the procedure had not happened to be covering all 3 hospitals, and if there had not happened to be relatively high rate of acute HCV infections leading to hospitalization, the cluster would have gone undetected, and practices that led to the infections would have continued. A more reliable approach—not dependent upon human factors— would involve automated surveillance to identify any acute hepatitis case within a specified period after endoscopy to prompt additional case review and evaluation. If this strategy is to be effective, automated infection surveillance across inpatient and outpatient facilities is necessary to allow linkage between the exposure and subsequent events. Ideally, genomics data would also be available to improve accuracy of case linkage, as was completed during the evaluation of several clusters of HBV and HCV infections following myocardial perfusion scans in a cardiology clinic. ^{Reference Tressler, Del Rosario and Kirby21}

Leveraging EHR data to predict rather than measure

Current IPC surveillance activities focus on infection detection. Retrospective measurement is helpful for identifying potential areas for quality improvement and for measuring the impact of new prevention programs. However, retrospective measurement is limited in its ability to prevent infections. Infection prevention based on prediction rather than detection has great potential in reducing HAI incidence. In recent years, technological advances in the EHR and data science have brought real-time predictive modeling into current clinical practice. Such tools have been used to predict clinical deterioration, sepsis onset, and more. ^{Reference Singh, Valley and Tang22–Reference Wong, Otles and Donnelly25} Predictive models could be used to identify patients at particularly high risk of developing HAIs and to target prevention strategies. ^{Reference Oh, Makar and Fusco26} Such tools could also be used to predict complicated or severe disease, enabling early intervention, and improving patient outcomes. ^{Reference Li, Oh, Young, Rao and Wiens27}