Study setting

This study was conducted at the University of Pittsburgh Medical Center-Presbyterian Hospital (UPMC), an adult medical and surgical tertiary-care hospital with 762 total beds, 150 critical care unit beds, >32,000 yearly inpatient admissions, and >400 solid organ transplants per year. The UPMC eRecord EHR system has >29,000 active users, including >5,000 physicians affiliated with UPMC, and it comprises >3.6 million unique electronic patient records. UPMC uses Cerner Millenium PowerChart (Cerner, Kansas City, MO) and EpicCare (Epic Systems, Madison WI) as the backbone of its inpatient and outpatient EHR systems, respectively. The University of Pittsburgh Institutional Review Board approved this study.

Characterization of retrospective outbreaks from 2011 to 2016

The subject of this study were outbreaks that occurred during 2011–2016 and that had been previously characterized using molecular typing and traditional epidemiologic methods to identify the transmission route. The routine infection prevention practice at the time was to notify the Microbial Genomic Epidemiology Laboratory (MiGEL) of suspected outbreaks caused by bacterial pathogens so that molecular subtyping could be performed. For each patient suspected of being included in the outbreak, the bacterial isolate was obtained from the clinical microbiology laboratory. For Clostridium difficile, which is diagnosed at our institution by culture-independent diagnostic testing, the nucleic acid amplification test–positive stool specimen was cultured for C. difficile. For identification of the common exposure responsible for individual outbreaks, our infection prevention team analyzed the health records of patients included in the outbreak to identify the responsible routes of transmission (eg, shared locations/staff, shared procedures/operations, or shared medications). Some outbreak investigations utilized environmental or device cultures to confirm the route of transmission. The transmission route defined by infection prevention was used as the gold standard for comparison with transmission routes identified by the data mining algorithm.

During the study period, our primary method for molecular characterization of bacterial isolates was pulsed-field gel electrophoresis (PFGE). To be considered part of the outbreak, patient isolates had to have 85% band similarity by PFGE. In 2016, WGS replaced PFGE. Based on our experience using WGS for hospital outbreaks, a cutoff of ≤20 single-nucleotide polymorphisms (SNPs) was used to define genetically related patient isolates.

Extraction and processing of EHR data for data mining

All inpatient, emergency room, and same-day surgery encounters between January 1, 2011, and December 31, 2016, were identified through an EHR data repository that accepts data from Cerner and EpiCare as full-text medical records and integrates information from central transcription, laboratory, pharmacy, finance, administrative, and other departmental databases.Reference Yount, Vries and Councill¹⁰ For each encounter, we obtained microbiology reports and charge transactions from the data repository. To maintain patient confidentiality, a research database was established in which each patient was assigned a study identification number (ID) using De-ID software (De-ID Data, Philadelphia, PA). Criteria for exemption from informed consent by the university’s institutional review board were met.

Charge transaction data are in a data repository as charge codes that contain patient account numbers, dates of service, cost centers, transaction codes, charge quantities, and charge amounts. Charge codes include any medication, procedure, location, or other billable item in our hospital and therefore encompass a large number of possible hospital exposures for each patient. Multiple charge codes can represent exposure to a single instrument; therefore, charge codes for key procedures (eg, endoscopic retrograde cholangiopancreatography [ERCP] and bronchoscopy) were collapsed into a single variable group that represented that exposure. For example, ERCP has 8 Current Procedure Terminology codes: 43260, ERCP diagnostic including collection of specimens; 43261, ERCP with biopsy; 43278, ERCP with ablation of tumors, etc. All were combined into a single variable called “ERCP,” although each charge code was also analyzed individually.

Data mining of the electronic health record (EHR)

The data mining program was designed using a case-control approach based upon the genotyping results using patient EHR data unrelated to the outbreaks as controls.Reference Miller, Chen, Sundermann, Marsh, Saul, Shutt, Pacey, Mustapha, Harrison and Dubrawski¹¹ Case patients were defined as those who had clinical isolates with the same strain by PFGE or WGS, as defined above. Controls were patients who were hospitalized during the 30 days before the case patients’ culture date and did not test positive for the genetically related bacterial species. Hospital exposures were then compared for cases and controls.

The data mining program was run on all 9 previously identified outbreaks identified by the infection prevention team at our institution during 2011–2016 to determine the sensitivity of the algorithm for identifying the correct transmission route. The transmission route was deemed to be correct if the route was ranked in the 3 most likely routes of transmission and/or had odds ratios >1 with significant P values. Preventable infections were calculated based upon a hypothetical 7- or 14-day infection prevention intervention from the date of the positive culture assuming that the data mining program had been running in real time and that effective interventions were enacted (eg, removal of contaminated equipment, disinfection of environment, and/or enhanced precautions). Outbreaks were deemed nonpreventable if there were only 2 isolates in the outbreak.

We scored possible common routes of transmission within an outbreak according to the formula:

$$S = a \ {\rm{ln}}\left( {a/r} \right){\rm{ }} + {\rm{ }}\left( {r - a} \right){\rm{ ln}}\left( {{\rm{1 }} - a/r} \right){\rm{ }}-a\ {\rm{ln}}\left( \gamma \right),$$

where a is the number of case patients exposed, r is the number of patients exposed overall (ie, case patients who are part of the outbreak and control patients who are not), and γ is a parameter that balances the positive and negative evidence (γ = 1e−4). For a given set of case patients, each patient can be said to have been infected through the hypothesized common route or by intermediate transmission (ie, via transmission from another case patient). If we take θ to be the unknown probability a patient becomes infected upon exposure to the hypothetical route and γ to be the probability a patient is infected by intermediate transmission (ie, by some other means such as patient-to-patient transmission), then the likelihood of observing a particular set of case patients is proportional to θ^b(1 − θ)^{r -b}γ ^-b, where b is the number of case patients infected by the route. We arrive at the formula S by maximizing this expression in θ, which occurs at θ = b/r, and b, which occurs at either 0 or a. Because b = 0 is a degenerate solution, it is disregarded. The final score is the log of this maximum likelihood.

The score above represents a nonnormalized log-likelihood. Because it is not normalized, it is suitable for ranking routes but is not comparable across time as the number of case patient changes. Therefore, we estimated an extreme value statistic as the probability a route would score at least as highly as its observed score under the assumption that the case patients were uniformly randomly sampled (the null hypothesis). This P value is estimated numerically using importance sampling from the observed data.

Researchers were initially blinded to the true routes of transmission in this analysis. However, it became clear during the development of the approach that this significantly reduced our ability to identify and correct data processing and modeling problems. For example, the charge codes for gastroscope procedures initially were not properly extracted and grouped from the EHR and therefore could not be identified in the analysis. On review, the correct charge codes for procedures using gastroscopes were grouped together as described above. The analysis was rerun, and the correct exposure route was identified.