EnABLe: An agent-based model to understand Listeria dynamics in food processing facilities

Detection of pathogens in food processing facilities by routine environmental monitoring (EM) is essential to reduce the risk of foodborne illness but is complicated by the complexity of equipment and environment surfaces. To optimize design of EM programs, we developed EnABLe (“Environmental monitoring with an Agent-Based Model of Listeria”), a detailed and customizable agent-based simulation of a built environment. EnABLe is presented here in a model system, tracing Listeria spp. (LS) (an indicator for conditions that allow the presence of the foodborne pathogen Listeria monocytogenes) on equipment and environment surfaces in a cold-smoked salmon facility. EnABLe was parameterized by existing literature and expert elicitation and validated with historical data. Simulations revealed different contamination dynamics and risks among equipment surfaces in terms of the presence, level and persistence of LS. Grouping of surfaces by their LS contamination dynamics identified connectivity and sanitary design as predictors of contamination, indicating that these features should be considered in the design of EM programs to detect LS. The EnABLe modeling approach is particularly timely for the frozen food industry, seeking science-based recommendations for EM, and may also be relevant to other complex environments where pathogen contamination presents risks for direct or indirect human exposure.


Introduction sub-processes
LS was introduced into the finished product processing room via three mechanisms: (i) carried into the room from rooms adjacent to the slicing room (Zone 4), during times of high traffic into the room (pz=10^Pert [-3.4,-2,-1.2,4]; expert opinion); (ii) introduced from food materials entering the room each day, i.e., cold-smoked salmon fillets being skinned, sliced and packaged (Rd=10^Pert[-7,-4,-1,4]; expert opinion); and (iii) introduced by unpredictable and undefined random events, which followed a Poisson process with the time to the next event assumed to follow an exponential distribution with mean pr=10^Pert [-3.4,-2,-1. 2,4] (expert opinion). Introduction from Zone 4 was restricted to floor patches and agents in proximity to doorways and was modeled by adding a LS load (Nz=10^Pert[0, 0.7, 2.0, 4.6] CFU; expert opinion) to the patch or agent and updating its concentration. Introduction from random events was possible at any patch (floor or ceiling) or agent in the slicing room and was modeled by adding a LS load (Nr=10^Pert[0, 0.7, 4.0, 5] CFU; expert opinion) to the patch or agent and updating its concentration. Finally, introduction from an incoming cold-smoked salmon fillet (100 g) to be processed was limited to a Zone 1 surface in the skinning, trimming or slicing area and was dependent on the concentration of LS (NR~Gamma[1.2, 0.19] CFU/g) 1 and the transfer coefficient (α=10^Normal[-0.28,0.2]) 2 . The model considered only these three modes of introduction which were identified and estimated through expert elicitation (described in Appendix II). These introduction modes expanded upon previous models that generally considered the presence of some reservoir in the processing facility as the source of LS capable of contaminating a food contact surface. It is recognized that there may be other potential modes of introduction, however they were not considered in this model.

Growth and survival sub-process
In the modeled environment, growth only applied to LS on patches and agents containing either moisture or visible water (and assumed presence of a sufficient amount of nutrients from residual food) at a particular time (i.e., during the current tick).Growth was modeled hourly according to a solution of the primary Verhulst logistic function, which describes growth as proportional to the present population and the available nutrients 3 : where Nt was the initial population in 1 cm of moisture on the surface at time t, CFU/cm 3 , K was the carrying capacity of the environment, 10 8 CFU/cm 3 4 , and µ was the maximum specific growth rate, h -1 , generated for each model iteration according to: where the generation time (GT, h) was uniformly distributed within the range from 8.4 to 24.2 (GT~Uniform[8.4, 24.2]) for 10°C (pH=5.6 and aw=1.0) 5-7 ; the modeled room maintains temperature at 10°C. This equation and growth rate matched the common practice of modeling microbial growth in foods and was thus considered as reasonable. In the modeled environment, this growth only applied to LS on patches and agents containing either moisture or visible water (and assumed presence of a sufficient amount of nutrients from residual food) at a particular time (i.e., during a particular simulated hour). LS did not grow in dry areas, but did survive; that is, in our model the LS population experienced no net change in the absence of water. Listeria has been shown to survive desiccation for extended periods of time in food processing environments and similar conditions 8,9 . It was assumed that there was no lag time and the physiological state of the cells (i.e., stressed, starved, exposed to sanitizer) was not considered to affect the growth rate.
These assumptions were reasonable and necessary as sufficient data to model lag phase or to include physiological state of LS on equipment surfaces were not available. Furthermore, the model did not consider formation of Listeria spp. biofilms on surfaces or attachments to equipment that become increasingly stronger with time or increasingly resistant to disinfectants.
It was assumed that cells are uniformly distributed on contaminated surfaces. These assumptions were made due to the high degree of uncertainty in our understanding of the involved processes and required model parameters and may present a model limitation. Several studies have concluded that persistent and presumed nonpersistent LM strains were equally susceptible to disinfectants 8,10 . The complexity of different biofilm growth rates and adhesion strengths has been acknowledged but excluded in a previous risk assessment model 11 . In other models, growth was not considered to occur on equipment surfaces at all 12,13 . Our use of site characteristics (i.e., cleanability and presence of water) and growth accounts for the possibility that Listeria persists on modeled surfaces and the environment.

Transmission sub-processes
LS transmission was modeled throughout the environmental patches and equipment according to hourly activities in the slicing room. Transmission occurred among patches, among agents, and between patches and agents, as described in the following sections.

Transmission among patches
LS was spread across the floor patches by foot traffic or within contiguous puddles of water. The presence of water was described as either absent (dry), moist, or visibly wet and was represented by a numerical code 0, 1, or 2, respectively, for each patch. Similarly, traffic patterns on the floor were observed and areas were classified as high, low, and negligible traffic. The traffic levels were assigned contact rates of 60 contacts/hr, 12 contacts/hr and 0.2 contacts/hr, respectively, based on observations. The water and traffic states of each patch were dynamic over the production shift depending on the activity in the slicing room each hour. For example, during cleaning, when water was observed to be used to spray down equipment prior to use of detergents, all agents and patches in the model were updated to be "visibly wet." Similarly, traffic was higher at the beginning and end of a shift compared to the middle of a production shift.
The probability of LS transmission ( ) via foot traffic at time t and patch j was based on the incidence rate ( ) of adjacent patches becoming contaminated from patch j due to the frequency of contacts associated with the traffic assigned to the patch j and its adjacent patches.
This process was modeled as: where was the probability that patch was selected from all patches containing traffic ( = 1/4593 = 0.0002); 1 was the fraction of patches adjacent to patch j that were in the same or higher traffic level; was the contact rate between the contaminated patch j and the adjacent patch given the traffic level at the contaminated patch j and was based on observations in the modeled room of the smoked-salmon facility of ℎ ℎ = 60/ ℎ/ℎ , = 12/ ℎ/ℎ , and their importance was evaluated in the sensitivity analysis.

Transmission among agents
Transmission onto and between equipment occurred based on connectivity with links, as previously described. The probability of contact (pij) between agents i and j was generalized based on their zone categories (Table S1). Similarly, the probability of LS transferred given contact between agents i and j (τij, the transfer coefficient) was also generalized by zone and assumed to be independent of the initial number of bacteria on the surface. No bacteria were lost during transfer events, thus conserving the overall mass. Within the surface population to be transferred, each bacterium was set to act independently such that the overall transfer was the result of a sum of independent Bernoulli trials, and modeled as a Binomial distribution, with the number of trials equal to the surface microbial population (CFU) and the probability of success equal to the transfer coefficient 12,13 . As supported by these references, this model setup matches the common practice in the modeling community.

Transmission between agents and patches
Transmission between the equipment agents and environment patches occurred via colocation, where co-location was defined as the presence of an agent (or agents) at different heights on the same patch coordinates. The two mechanisms modeled were: (i) condensation falling from the ceiling to either an agent or the floor below per hour, which was assumed to occur with probability pc~Uniform[0.01,0.05] and (ii) food falling from a Zone 1 equipment to a floor patch below during production (pf~Uniform[0.20,0.40], observed by the food safety manager in the processing room). The values of pc were assumed as no data were available, however their parameter values were intentionally set to wide ranges to test their importance in the sensitivity analysis. The model considered only these two modes of transmission which seemed to be the most relevant to LS transmission between agents and patches.

Removal sub-process
In addition to contamination removal by food products being processed, LS was removed from equipment and the environment due to mechanical elimination and disinfection during routine cleaning and sanitation. It was assumed that all bacteria cells on the same surface underwent the same cleaning and sanitation independently, that the overall reduction was immediate, and the resulting population was the sum of independent Bernoulli trials, and modeled as a Binomial distribution, with the number of trials equal to the surface microbial population (CFU) and the probability of success equal to ten raised to the expected daily log reduction, ηd 12,13 . The expected log reduction is sampled each day from a Pert distribution (ηd ~Pert[-8, -6,-1.5, 4]) 13 . Therefore, this assumption matches the practices accepted in the modeling community. The model is capable of modeling the presence of sanitizer in the production environment at both lethal and inhibitory concentrations, with the cutoff and importance dependent upon each sanitizer and the frequency with which its concentration is monitored. In the model, areas containing inhibiting concentrations of sanitizer included areas around doorways, where door-foamers or powders may be used, and around drains, where pooling of sanitizer occurs after sanitation. Powdered quaternary ammonia (1.25 g/25 cm 2 ), commonly used around doorways in ready-to-eat food production facilities, has been previously shown to have inhibitory effect on microbial growth but no significant reduction on LM 15 .
Therefore, as in dry areas, in areas of inhibitory sanitizer levels, LS was not able to grow but could survive.

Cleanability of the equipment and environment
It is well-understood that certain food processing equipment is unable to be effectively cleaned due to its design or the presence of holes or cracks, meaning even when detergent and sanitizer are applied Listeria, if present, will remain (along with water and organic matter, in so called niches or harborage sites) [16][17][18] . Often, hard-to-clean (referred to as "uncleanable") equipment is disassembled and cleaned as part of routine cleaning and sanitation or corrective actions following the detection of contamination. Thus, EnABLe agents were set to be either cleanable or uncleanable (Table S2) based on the understanding of a site's unique properties and properties of sites that likely represent niches 19 ; however, this status can be changed by the user.
For cleanable sites, there was still a probability (γ) that cleaning was not properly executed at the end of the shift, resulting in no reduction of LS on the surface. This random probability of changing the cleanable to uncleanable status was set at 0.01 based on the recognition that random events, such as human error (i.e., no disassembly of equipment on a given day), could result in unsuccessful cleaning of a site. The value for γ was investigated through scenario analysis and showed no differences when comparing baseline model conclusions to model conclusions for γ=0.1 and 0.001.

Environmental monitoring (EM) sub-process
During EM sampling, it was assumed that contamination was homogeneously distributed on the agent or patch surface and that the probability of detection of LS was dependent on the concentration on the surface. A 10% chance of false negative was assumed for concentrations between 0-10 CFU/cm 2 ; a 1% chance of false negative was assumed for concentrations between 10-100 CFU/cm 2 ; and, it was assumed that the chance of false negative was negligible for concentrations greater than 100 CFU/cm 2 . False positives were not considered possible due to detection of LS over LM and the relatively high specificity for advances in industry-utilized detection methods, especially nucleic acid-based methods 20 . These assumptions were based on expert opinion and the chance of a false negative being dependent on concentration was similarly modeled by Gallagher et al. 11 . The sensitivity of the model to the cut-off values for each chance of false negative was evaluated by scenario analysis and was not shown to impact model conclusions. The two schemes for evaluating the chance of false negative assumption were 0-20 CFU/cm 2 , 20-200 CFU/cm 2 , and ≥200 CFU/cm 2 and 0-5 CFU/cm 2 , 5-50 CFU/cm 2 , and ≥50 CFU/cm 2 for 10%, 1% and negligible probability of false negative, respectively.

Expert elicitation
Five experts from academia (2) and industry (3)   Your identity will be kept confidential.
Q1. 2 We are developing a model to study the behavior of microbial contamination (i.e. Listeria spp.) in the food processing environment. Your participation will help to elicit information regarding (1) how Listeria spp. may be introduced to and remain present in a finished product room of a food processing facility, and (2) Table 1 to give the probability that each occurs and provide your level of confidence in this answer.

Q3.2 Objects that move into and out of the room (i.e. trolley, cart, product bins)
In the event these objects enter into the finished product room, it is possible that the object: • brings Listeria spp. into the room, and • contacts another item in the finished product room, and • transfers Listeria upon contact with another item

Q3.3 Employee's hands
Each time an employee enters the finished product room, it is possible that the employee: • has Listeria on their hands, and • does not properly wash and disinfect their hands upon entry into the finished product room, and • contaminates their gloves with their hands

Q3.4 Employee's work-assigned footwear
Each time an employee enters the finished product room, it is possible that the employee: • has Listeria on their work-assigned footwear, and • does not properly scrub or cover footwear Q3.5 Food entering the finished product room Each time food enters the finished product room to be processed, it is possible that the food item: • is contaminated with Listeria, and • transfers Listeria to the first surface it contacts, and • transfers Listeria to subsequent surfaces it contacts Q3.6 Unpredictable event (i.e. roof leak, maintenance, drain backs up) During a shift, there may be events that: • cause interruptions or unplanned stops in production, or • bring visitors or additional employees into the room, or • increase the likely presence of Listeria in the room Q3.7 Taking all of these steps into account, for each scenario give the category that describes the probability Listeria spp. is introduced into the environment or equipment of the finished product room during different times and provide your overall level of confidence  Table 2, and then provide your confidence in these responses. For each scenario below, suppose there is a mode of contact (indirect or direct) between ITEM A and ITEM B and that ITEM A is contaminated (Listeria is evenly distributed on its surface). What is the probability (%) that transfer of Listeria spp. from ITEM A (contaminated) to ITEM B (uncontaminated) occurs within an hour of production? Write the range of probability in the corresponding empty boxes (without 'xx').
Q5.2 At the slicer, Employee 1 feeds product onto the slicer in belt. After it is automatically sliced, Employee 2 takes the product from the slicer out belt and moves it to a conveyor belt. There is a drain below the slicer. What is the probability with which transfer of Listeria spp. from ITEM A (contaminated) to ITEM B (uncontaminated) occurs within an hour of production? Write the probability range (i.e. 20-30%) in the corresponding empty boxes (without 'xx'). Q20 This is the end of the survey. Click back (<<) to review/change your responses. Click next (>>) if you are finished and are ready to SUBMIT your responses. You will not be able to make changes or access the survey once submitted.       -top  scale-table-scale  scale-table-scalepanel  scale-table-framework  scale-table-shelf  scale-table-top  scale-table-scale  scale-table-scalepanel  scale-table-framework  scale-table-shelf  scale-table-top  scale-table-scale  scale-table-scalepanel  scale-table-framework  scale-table-shelf  scale-table-top  scale-table-scale  scale-table-scalepanel  scale-table-framework  scale-table-shelf  scale-table-top  scale-table-scale  scale-table-scalepanel  scale-table-framework  scale-table-shelf  scale-table-scale  scale-table-scalepanel Table S5. Agent cluster analysis results according to contamination outcomes  -frame  cutting-table-top  cutting-table-leg  cutting-table-underside  cutting-table-frame  cutting-table-leg  cutting-table-underside  cutting-table-frame  cutting-table-leg  cutting-table-underside  cutting-table-frame  cutting-table-leg   drain  drain  employee  employee  employee  employee  employee  employee  cutting-table-