Attribution of Illnesses Transmitted by Food and Water to Comprehensive Transmission Pathways Using Structured Expert Judgment, United States

Illnesses transmitted by food and water cause a major disease burden in the United States despite advancements in food safety, water treatment, and sanitation. We report estimates from a structured expert judgment study using 48 experts who applied Cooke’s classical model of the proportion of disease attributable to 5 major transmission pathways (foodborne, waterborne, person-to-person, animal contact, and environmental) and 6 subpathways (food handler–related, under foodborne; recreational, drinking, and nonrecreational/nondrinking, under waterborne; and presumed person-to-person-associated and presumed animal contact-associated, under environmental). Estimates for 33 pathogens were elicited, including bacteria such as Salmonella enterica, Campylobacter spp., Legionella spp., and Pseudomonas spp.; protozoa such as Acanthamoeba spp., Cyclospora cayetanensis, and Naegleria fowleri; and viruses such as norovirus, rotavirus, and hepatitis A virus. The results highlight the importance of multiple pathways in the transmission of the included pathogens and can be used to guide prioritization of public health interventions.

The Centers for Disease Control and Prevention (CDC) works to control and prevent illness caused by foodborne and waterborne pathogens in the United States.To accomplish this, CDC supports states and territories in tracking disease, detects and responds to outbreaks, and uses surveillance and sentinel site data to estimate the burden of these diseases in the United States.To inform this work, we implemented an SEJ study using Cooke's classical model to estimate the proportion of domestically acquired illnesses for 33 pathogens transmitted through food and water that can be attributed to each of 5 major transmission pathways and 6 subpathways (12).

Methods
The process was divided into 3 stages: preparation, elicitation, and postelicitation (11).These stages are detailed in the following sections.

Preparation Selection of Pathogens
We included all pathogens transmitted commonly through food or water that were examined by Scallan et al. (2) and Collier et al. (13) except those for which the only syndrome of interest was considered to have >95% foodborne transmission (e.g., Listeria monocytogenes, Clostridium botulinum); we added 3 free-living amoebae (2,13).For some pathogens, subdivisions into categories by serotype, patient age, or clinical manifestations of interest were included because transmission pathways were assumed to be different.For example, for Salmonella, the 5 most common serotypes were included along with 2 groups of rarer serotypes based on a ranking of their coefficients of variation (CVs) calculated from the patients' ages, sexes, states of residence, and the year and month specimens were obtained (group 1, lowest CVs; group 2, highest CVs) as described by Boore et al. (14).This compilation resulted in a total of 33 pathogens and 47 target questions, or categories, for estimation.The 47 target questions were grouped into 15 panels on the basis of similarities between pathogen microbiology and ecology (Table 1).

Transmission Pathway Definitions
We used definitions for 5 major pathways that were mutually exclusive and comprehensive (i.e., covering 100% of transmission modes) and that reflect those used by CDC for outbreak surveillance (15,16; Tables 2, 3).We defined 3 mutually exclusive waterborne subpathways (recreational water, drinking water, and nonrecreational nondrinking water) that were comprehensive (i.e., all waterborne pathway transmission fell into 1 of the 3 subpathways).We also defined and elicited 1 foodborne (food handler-related) and 2 environmental (presumed animal associated, presumed person-to-person) subpathways that accounted for only a portion of transmission within their main pathway.We calculated the unelicited proportion remaining of their respective main pathways during analysis and assigned it to the subpathways other foodborne and other environmental.For all transmission pathways, we defined the point of attribution as the point of exposure (i.e., the event during which a person ingested, or was otherwise exposed to, the pathogen).

Expert Identification and Selection
We identified 182 experts representing a range of scientific backgrounds (e.g., epidemiologists, laboratory scientists, and environmental engineers from government, academia, nongovernmental organizations,   and industry) on the basis of publication records, experience, expertise, or previous participation in source attribution studies.We contacted the experts directly and invited them to apply for participation (Figure 1).Fifty-eight returned a curriculum vitae and publication record and completed a questionnaire about their professional interest, knowledge, and experience for each of the 33 pathogens using a 4-level Likert scale (high, medium, low, or none) by the requested deadline.We asked experts to suggest additional experts to be considered; the 3 who were suggested were also invited.

Assignment to Panels
We evaluated expert applications based on area of expertise, education, work history, professional interest, experience, and knowledge of the individual pathogens in this study.Publication record was not used to determine eligibility because it could have led to elimination of qualified experts who do not publish frequently.We used maximum bipartite matching in R version 3.3.1 with the igraph package version 1.0.1 to assign experts to panels based on their curricula vitae, publication records, and questionnaire responses (17,18).Final assignment ensured that experts were not on pathogen panels for which they reported none or low experience.Individual experts were on panels for <15 pathogens (Appendix 1, https://wwwnc.cdc.gov/EID/article/27/1/20-0316-App1.pdf).

Calibration Questions
The study administrators used unpublished data to develop calibration questions (Appendix 2, https://wwwnc.cdc.gov/EID/article/27/1/20-0316-App2.pdf).We developed 14 questions to evaluate the experts' statistical accuracy and informativeness by probing the experts' ability to provide reliable estimates under uncertainty.The subject domain of the questions aimed to represent expertise in public health surveillance of foodborne and waterborne diseases, food consumption patterns in the United States, and human exposure and occurrence data about pathogens in food, water, and the environment.

Target Questions
Target questions asked the proportion of illnesses transmitted through the 5 major pathways and 6 subpathways for all study pathogens.Study administrators blocked transmission pathways and subpathways for some pathogens based on their microbiology and ecology (Table 4).We created individualized Microsoft Excel version 14.

Dry Run Exercise
We conducted a dry run exercise using video web conferencing to assess calibration questions, target question answer sheets, and expert training materials for completeness, clarity, and ease of use.Six persons from academia, state health departments, and CDC participated in this trial exercise, but not in the formal elicitation itself.We modified the elicitation materials based on feedback from this exercise.

Expert Orientation
Before the formal elicitation, experts attended a training webinar to learn definitions of transmission pathways, subpathways, and point of attribution.To ensure common understanding of the definitions, experts completed a 20-question review of knowledge after the webinar (Appendix 4, https://wwwnc.cdc.gov/EID/article/27/1/20-0316-App4.pdf).
We provided a background document summarizing current surveillance data, when available, and relevant research findings for each pathogen.The document contained links to selected research articles.Experts were encouraged to use any data they felt were informative to make their estimates; they were not limited to only this document.

Elicitation
For the formal elicitation, 48 experts representing a wide range of professional and scientific backgrounds participated at a 2-day, in-person workshop in May 2017.During the workshop, experts participated in a 2-hour information session on probabilistic methods and providing estimates under uncertainty.

Calibration Questions
Experts were not expected to know true values precisely and provided low (5th percentile), median (50th percentile), and high (95th percentile) estimates to represent their uncertainty on the answers provided to the calibration questions.Experts were not allowed access to any additional resources while answering the calibration questions and, after they had they had finished, they could not return to this section to change their responses.

Target Questions
After completion of the calibration questions, experts provided 5th, 50th, and 95th percentile estimates for the proportion of domestically acquired illnesses that are transmitted through each major pathway and subpathway annually for each pathogen and target question in each panel to which they were assigned.
The experts were also asked to indicate if they did not agree with the pathways blocked by study administrators.One pathway, person-to-person transmission for Legionella spp., was unblocked based on this feedback, and experts provided this estimate with the others at the in-person elicitation.Experts could access resources and discuss them with colleagues, if desired.However, we emphasized that the final estimates should represent the expert's individual responses, not a group consensus.

Re-Elicitation
After the in-person elicitation was completed, we determined that re-elicitation for some pathogens was necessary.More granular detail was needed beyond the single estimate for Pseudomonas, so estimates were re-elicited for otitis externa, septicemia, and pneumonia.Based on feedback we received during the elicitation, we re-elicited estimates for non-acute gastroin-testinal infections (non-AGI) for nontoxigenic Vibrio cholerae, V. parahaemolyticus, V. vulnificus, and V. alginolyticus.Experts were provided with feedback with updated surveillance data and given the opportunity to adjust their original estimates if new data led them to reconsider their previous estimates (Figure 1).The re-elicitations were completed through follow-up emails and web conferences.

Data Analysis
We analyzed data using EXCALIBUR ( 19).We combined all individual expert assessments by linear pooling into a single uncertainty assessment for each target question (11).For equal-based weighting, all experts' assessments contributed to the combined uncertainty assessment evenly.We computed performance-based weighting by combining the statistical accuracy and information scores of experts in each panel.The weighted combination of experts is referred to as the decision maker.We used the item weight decision maker because this calculates and applies weights per individual target question rather than for all questions an expert answered.We performed optimization to determine the threshold by which an expert's responses would be included in the final estimate or not.This was done separately per expert for each panel, based on each expert's statistical accuracy score (12).We performed a subgroup analysis to determine whether separate schools of thought existed as a result of experts' self-identified background (categorized as mainly foodborne, mainly waterborne, or both).This analysis was completed by 2 independent reviewers who analyzed EXCALIBUR panel outputs for each target question to determine whether wide divergence existed among individual responses.
We normalized random samples from the weighted distributions for major transmission pathways and waterborne subpathways such that on each sample the values across pathways summed to 1.This process was done by resampling the cumulative distribution functions generated by EXCALI-BUR 5,000 times in R version 3.4.3for each pathogen, while dividing all sampled values by the sum of their values per iteration.Point estimates and 95% uncertainty intervals (UIs) for each target question and pathway were produced.We performed robustness analysis and out-of-sample validation to assess the performance of the method and to evaluate the effect of individual experts and individual calibration questions on the final distribution (Appendix 5, https://wwwnc.cdc.gov/EID/article/27/1/20-0316-App5.pdf)(12).

Knowledge Review
The 20 questions were designed to be challenging, to emphasize application of the study definitions, and to represent scenarios at the boundaries among different transmission pathways.For 17 (85%) questions, >75% of participants answered with the correct major pathway, and of these questions, 13 (76%) were answered with the correct subpathway as well (Appendix 4).

Major and Subpathway Results
Table 4 and Figures 2 and 3 show the proportion and UI of domestically acquired illnesses attributed to the 5 major transmission pathways; Tables 5 and 6 show the subpathway results.For all panels, a satisfactory number of accurate and informative experts were included.Differing schools of thought based on experts' backgrounds were not identified (Appendix 5).

Bacteria
Most of the pathogens in this study were bacteria; they encompassed 35

Viruses
Most transmission for astrovirus, norovirus, rotavirus, and sapovirus was attributed to the person-toperson pathway.Hepatitis A virus was estimated to have the highest proportion of illness transmitted by the foodborne pathway at 42% (UI 9%-78%).Of this, 48% (UI 2%-93%) was considered food handler related.Of foodborne transmission, 50%-71% was estimated to be food handler related for astrovirus, norovirus, and sapovirus.For all viruses, 67%-88% of environmental transmission was attributed to the subpathway of presumed person-toperson transmission.

Discussion
This study presents a novel method for estimating the proportion of illnesses from pathogens transmitted commonly by food and water in the United States  non-O157, other diarrheagenic E. coli, nontyphoidal S. enterica, M. bovis, Shigella spp., Y. enterocolitica, C. cayetanensis, T. gondii, astrovirus, rotavirus, sapovirus, and hepatitis A virus.These differences could be the result of changes in data availability or analytic methods.For example, previous US foodborne illness estimates used data from surveillance, risk factor studies, and literature review (2).Based on available data for S. enterica (a case-control study of sporadic illness and unpublished outbreak data [2,25]), a study used an estimate of 94% foodborne transmission, notably higher than this study's estimate of 66% (UI 48%-81%).Estimates more similar to the current study were reported in SEJ studies in the Netherlands (55%), Canada (63%), and Australia (71%) (21,22); these studies examined attribution to similar major pathways to those included in this study versus foodborne transmission only.Our estimates of foodborne transmission of astrovirus (15%), rotavirus (5%), and sapovirus (13%) are much higher than the estimate of <1% for each in an earlier study (2); reports of foodborne outbreaks caused by these viruses in CDC's outbreak surveillance systems informed our estimates.Reporting of enteric disease outbreaks transmitted by nonfoodborne routes has improved, and experts probably used these new data to inform their estimates ( 26).
This study provides noteworthy estimates for the food handler-related subpathway.For hepatitis A, both the World Health Organization and this study estimate 42% foodborne transmission, of which this study estimated 48% (UI 2%-93%) to be food  handler-related (20).However, this study was conducted before widespread awareness of a massive increase in person-to-person transmission in the United States (27).Previous estimates of foodborne transmission were 11% in the Netherlands and 7% in the United States (2,21).The use of different pathway definitions, points of attribution, and inclusion of travel-related illness in these other studies might have contributed to these differences (21,28).For norovirus, 71% (UI 29%-99%) of foodborne transmission in our study was attributed to the food handler subpathway, which is supported by studies of outbreaks in the United States (29,30).
For the waterborne transmission pathway, attribution in the context of the other pathways has not been done before in the United States.Furthermore, these estimates include subpathway estimates and non-gastroenteritis clinical outcomes.For bacterial pathogens, the estimates suggest that the proportion of illnesses linked to water is higher than previously appreciated.The estimates for waterborne bacterial pathogens were associated with high rates of illness and death, including nontuberculous Mycobacterium spp., Pseudomonas spp., and Legionella spp.Of note, neither Giardia spp.nor Cryptosporidium spp., parasites traditionally understood to be waterborne, were assessed as predominantly waterborne; instead, person-to-person and animal contact, particularly for Cryptosporidium, were key pathways.For the free-living amebae Acanthamoeba spp., B. mandrillaris, and N. fowleri, limited data are available on exact exposures associated with these rare illnesses (31,32).The proportion of viral pathogens transmitted by water was estimated to be relatively low (6%-8%), although for norovirus this represents a substantial proportion of estimated annual waterborne disease illnesses (32).This study also provides estimates for 3 waterborne disease subpathways.Of note is the proportion of otitis externa infections caused by Pseudomonas spp. that were attributed to recreational water exposure, and the combined contribution of drinking and nonrecreational, nondrinking water exposures to nongastroenteritis outcomes of Pseudomonas spp.(excluding otitis externa), nontuberculous Mycobacterium spp., and Legionella spp.CDC has used results from this SEJ to help estimate that 7.2 million waterborne illnesses occur from 17 pathogens annually, including 600,000 emergency department visits, 120,000 hospitalizations, and 7,000 deaths, incurring $3.2 billion (2014 US dollars) in direct healthcare costs (33).Whereas the primary focus of this SEJ study was illnesses transmitted commonly by food and water, including person-to-person, animal contact, and environmental transmission was integral to the study and led to notable findings.For example, this study estimated animal contact transmission of STEC O157 at 12% (UI 3%-25%) and of STEC non-O157 at 21% (UI 2%-46%).Previous US animal contact estimates, which were based on a FoodNet case-control study and outbreak surveillance data, estimated STEC O157 at 6% and STEC non-O157 at 8% (24).This discrepancy may be the result of differences in pathway definitions and the inclusion of additional data.
As with other SEJ studies, this study is subject to limitations that can affect the interpretation of results.Estimates for many pathogens had wide UIs, highlighting areas in which data gaps remain and further investment into public health surveillance and research may be warranted.More detailed attribution, such as by food category, was beyond the scope of this study.This study considered attribution at a national level and does not represent the geographic variability that exists for some pathogens.Experts provided estimates considering data available during the elicitation session, but infectious disease epidemiology can change rapidly, so these results may not reflect current transmission patterns.New information should be considered when applying these estimates (e.g., for disease burden calculations).Expert fatigue may have been a factor for participants who were asked to provide estimates for a large number of pathogens.For intervention and policy-making purposes, these results should be considered in context with results from other data-driven approaches, such as those done by the Interagency Food Safety Analytics Collaboration and for the Model Aquatic Health Code (34,35).
In conclusion, our findings provide a balanced understanding of multiple routes of transmission for 33 pathogens.This information can be used to support appropriate targeting of resources to prevent infections transmitted by all pathways and to invest in research and surveillance.

Figure 1 .
Figure 1.Expert selection process for study of attribution of illnesses transmitted by food and water to comprehensive transmission pathways using structured expert judgment, United States, 2017.

Figure 2 .
Figure 2. Source attribution results for major transmission pathways of bacteria in study of attribution of illnesses transmitted by food and water to comprehensive transmission pathways using structured expert judgment, United States, 2017.

Table 2 .
Major transmission pathway definitions, structured expert judgment, United States, 2017 Major transmission pathways Description FoodborneTransmission occurs through eating food.Contamination can originate anywhere in the food production chain from primary production, to retail, and then to the home or restaurant.This pathway applies to all nonwater beverages and items ingested by humans as food (e.g., including raw milk and excluding items consumed for medicinal purposes).

Table 5 .
Source attribution results for foodborne and environmental transmission subpathways, structured expert judgment, United States, 2017* *Blocked indicates pathways blocked by study administrators.AGI, acute gastrointestinal disease; STEC, Shiga toxin-producing Escherichia coli.†Clinicalmanifestations of interest for initial elicitation were bacteremia and wound infections.

Table 7 .
Comparison of proportion of illnesses attributed to foodborne transmission from this and earlier studies* *NE, not estimated; SEJ, structured expert judgment; STEC, Shiga toxin-producing Escherichia coli.