Food Control 29 (2013) 221e225 Contents lists available at SciVerse ScienceDirect Food Control journal homepage: www.elsevier.com/locate/foodcont Relating the biotracing concept to practices in food safety Anca Ioana Nicolau a, *, Gary C. Barker b, Iuliana Aprodu a, Martin Wagner c a Faculty of Food Science and Engineering, “Dunarea de Jos” University of Galati, 111 Domneasca Street, 800201 Galati, Romania Institute of Food Research, Norwich Research Park, Norwich, United Kingdom c Institute for Milk Hygiene, Department for Farm Animals and Veterinary Public Health, Vienna, Austria b a r t i c l e i n f o a b s t r a c t Article history: Received 23 December 2011 Received in revised form 27 April 2012 Accepted 9 May 2012 Biotracing, a new method developed to assist with food chain management and to control food safety, is presented to highlight practical considerations, including logistic issues for its implementation. The main differences between traceability and biotracing and between predictive microbiology and biotracing are explained and examples of situations in which biotracing could be of real help are listed (foodborne outbreaks, liability issues, HACCP, risk evaluation and decision making, education and training). Indications on how to access and interrogate two prototype models, called SimpleTrace and SimpleMatch, as well as some other Bayesian networks, are given to encourage using of biotracing, while operational biotracing is illustrated by an agent based model called AgentChain. The main types of inferences, which point to sources that generate potential problems within a particular food chain, are revealed. Biotracing is strongly recommended for introduction into continuous operations that include in line data collection, and can be operated, alongside existing safety systems, without additional burden. Ó 2012 Elsevier Ltd. All rights reserved. Keywords: Biotracing Food safety Liability HACCP Risk analysis Outbreaks Inferences 1. Introduction Defined as “the ability to use down stream information to point to materials, processes or actions within a particular food chain that can be identified as the source of undesirable agents” (Barker, Gomez, & Smid, 2009) or shorter “the ability to identify the sources of microbial contamination in a food chain” (Smid, Swart, Havelaar, & Pielaat, 2011), biotracing/biotraceability is a new research area that is fast developing and will soon become an essential tool to assist food safety specialists in quantitative assessment of microbial hazards. Although biotracing is strongly connected with risk assessment, and takes a comparable system wide view of contamination hazards, it concentrates on the identification of sources and origins rather than on the quantification of end effects (see e.g. www.biotracer.org). In this respect biotracing is a substantial addition to the food chain management process as it supports decision making and targeted actions. Source tracing has recently been identified as a powerful addition in many areas of microbiological investigations (Dance, 2008). Although predictive microbiology is an important element of biotracing (Hoorfar, Wagner, Jordan, Boquin, & Skiby, 2011), predictive microbiology and biotraceability are different concepts (Table 1). Both use mathematical models for predictions and hence imply a detailed knowledge of the microbial ecology and * Corresponding author. Tel.: þ40 (0)336 130 177; fax: þ40 004 0236 460 165. E-mail address: anca.nicolau@ugal.ro (A.I. Nicolau). 0956-7135/$ e see front matter Ó 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.foodcont.2012.05.020 environment, but predictive microbiology estimates the growth, survival or death of microorganisms (Ross & McMeekin, 2003), while biotracing expresses posterior beliefs about possible microbial origins or causes of growth (sources) (Smid, Verloo, Barker, & Havelaar, 2010). In this respect it is possible to consider predictive microbiology and biotracing as two sides of a coin, one causal and one inferential, that relate to the population dynamics for bacterial populations (and other agents) within a well defined food chain system. When compared to the traditional tracing process (Table 1), which mainly involves monitoring, record keeping and summarizing permanent information on a label or tag, the biotracing process is clearly distinct and complementary: biotracing establishes an information set and a belief system that can support the identification or prioritization of potential initiation points or processes based on uncertain detection of hazards or hazard markers (Barker et al., 2009). In addition to communication of the biotracing concept, to producers, traders and decision-makers, this paper adds to the development of biotraceability as a discipline and emphasizes its potential as investigational and educational tool. 2. Biotracing models In practice biotracing is highly multidisciplinary and involves a strong integration of food chain and microbiological expertise with mathematical modeling. As well as recently published models 222 A.I. Nicolau et al. / Food Control 29 (2013) 221e225 Table 1 Differences between traceability, biotraceability and predictive microbiology. Traceability Biotraceability Predictive modeling Relationship to food chain Capabilities Compulsory along the food chain Able to identify origins, to establish the history of a food product (showing from where the food comes) and track the process Complementary Able to predict micro-organism development (estimating cell numbers) and indicate potential problems Supported by Codes on labels or tags Data collected by Monitoring and recording Data format Non-standardized, although increasing uniformity Relational Complementary Able to identify, assess and rank vulnerabilities in the food chain (revealing where did the problem arise and what was the most probable cause Surveillance, information technology and statistical inference Multiple information streams from biology, process and historical record Probabilistic formulation of evidence, Bayesian interpretation of beliefs Inferential Character that describe biotracing for Salmonella in the pork chain (Smid et al., 2011), and Staphylococcus aureus in the milk chain (Barker & Gomez-Tome, 2011) other models currently being developed include Salmonella in the feed chain and Listeria in cheese. Additionally several models that partially express biotracing ideas have been published including source tracing for foodborne norovirus (Verhoef et al., 2010) and for Listeria in Gorgonzola cheese (Lomonaco et al., 2009). Operational biotracing involves significant efforts and developments in techniques for sampling, analysis and diagnostic modeling. These advances have resulted from massively improved rapid (molecular) methods for identification and enumeration of bacteria, e.g. Fourier transform infrared microscopy (Carlos, Maretto, Poppi, Sato, & Ottoboni, 2011), and molecular typing methods (Foley, Lynne, & Nayak, 2009). Additionally biotracing often includes an emerging use of advanced (Bayesian) computational methods for inference (Jarman et al., 2008). Two simple computational models, called SimpleTrace and SimpleMatch, have been used to illustrate the biotracing concept and to identify distinct patterns and motifs that are important in food chain biotracing. SimpleTrace is based on Bayesian inference within a well defined chain of uncertain events, while SimpleMatch is based on the statistics of type matching within a system of agent types and uncertain sources (See Table 2). These models, and some other Bayesian networks, are fully documented and can be accessed and interrogated on a development website hosted by HUGIN BIOTRACER (http://bbn.ifr.ac.uk). In very simple terms each of these models allows a user to use an observation (down stream information) to establish belief (posterior probability) about whether particular sources were significant in leading to the evidence. Both these schemes rely on Bayesian inference and both operate, consistently, even when information is incomplete or uncertain. In practical situations the inversion of causal reasoning, SimpleTrace, is often the easiest scheme to Historical data supplies and mathematical models Microbiological experiments and analysis Standardized and organized as searchable databases Causal interpret but, with an increasing ability to type and subtype harmful agents (particularly bacteria) the matching process may become the dominant biotracing process in the future. These two models illustrate an important property of operational biotracing; they can be used, consistently, with sequential pieces of evidence without ongoing development. However, in practice, it is often straight forward to employ a data stream both as a continuous source of evidence to identify problematic sources and also as support for a developing model that either increases confidence in inference or accommodates slow trends in the data supply. Computer learning techniques are compatible with the Bayesian network structures used to implement biotracing. The complex nature of decisions that surround food chain management, and hence operational biotracing, are illustrated by an agent based model, called AgentChain (See Table 2), that can be accessed and interrogated online (http://bbn.ifr.ac.uk). The AgentChain system includes spatially distinct chains, stochastic sources of agents, operational hurdles, batch partition, unit dispersion, agent population kinetics, unit consumption events, doseeresponse behavior and case identifications. The system is augmented by agent typing, record keeping and event monitors. This model shows that the sampling frequency and the decision taken on actionable evidence contribute to overall efficiency and safety of a process but that overall optimization is a complex task that could be aided by biotracing information. 3. Sources and inferences In practice biotracing often begins with the identification and definition of sources that characterize potential problems within a particular food chain. The identification of sources is the first step to the realization of a biotracing question e what is the most likely place that an observed problem has arisen. Within the context of biotracing a ‘source’ is quite loosely defined. Whilst it is clear that in Table 2 Simple models illustrating biotracing operations. Model Model type Location Biotracing principal illustrated Example application SimpleTrace Bayesian network 1 Inverting chain probabilities SimpleMatch Bayesian Network 2 Statistics of type matches AgentChain Agent based simulation 3 Complex optimization Separation of milk chain sources based on coupled measurements of enzymes and bacteria Elimination of food handler sources based on type mismatches Designing sample frequency to achieve an optimum chain efficiency and safety 1 e http://bbn.ifr.ac.uk/btmodeller/index.php/SimpleTrace. 2 e http://bbn.ifr.ac.uk/btmodeller/index.php/SimpleMatch. 3 e http://bbn.ifr.ac.uk/gary/biotracer/agentchain/Chain6.html. A.I. Nicolau et al. / Food Control 29 (2013) 221e225 some cases individual materials or ingredients might carry exceptional bacterial loads and therefore act as sources of contamination problems it is equally possible that manufacturing processes that are designed to reduce load might not operate fully and so also manifest as a contamination sources. In this way the biotracing sources, to some extent, parallel the initial loads and reduction steps included in a safety assessment that involves a food safety objective (Gorris, 2005). For biotracing sources can be effectively grouped together, e.g. all the post process contamination opportunities might initially be collected as a single source, and should be exclusive; but not necessarily exhaustive (that is each source can be identified only once in the list of sources but there might be unknown sources that are not identified in the list). In practice the inference step is a process of assigning probabilities to the sources based on evaluation of observed evidence. In the majority of normal operating conditions the ‘posterior’ probabilities are not significantly different from those that represented initial (prior) beliefs and are included in the biotracing model. However, exceptionally, evidence arises that expresses increased probabilities for one (or more) sources. The significant change in probability is a ‘biotrace’. In practice, largely based on legal precedent, the posterior probability is converted to another measure, called the likelihood ratio, as the basis for decision making (Taroni, Aitken, Garbolino, & Biedermann, 2006). The identification of sources and evaluation of posterior probabilities is sufficient to answer questions that relate to the sources of contamination hazards. Typical questions include: “for a set of possible sources, which is the most likely cause of actionable evidence?”; “for a particular source is the likelihood of initiating a hazard acceptable?”; “does evidence support the existence of unidentified sources?” and “for a particular source does evidence indicate fluctuations or a significant trend?”. In initial biotracing applications the first question has dominated. 4. Applications of biotracing Biotracing has a multitude of potential applications. Some situations in which biotracing could help are presented below: 4.1. Foodborne outbreaks Public health investigations often trace an outbreak of foodborne illness to a particular food or to a specific geographical source. The trace is often based on consumption patterns alone but sometimes is based on a matching process for the pathogen of concern. Although this trace is sufficient to prevent escalation or immediate repeat it is often not enough to identify a particular ingredient or process failure as the source of problems. In this respect the trace is often the precursor for much larger and more expensive investigations to seek the root cause of the problem. Food chain biotracing can go further by including evidence from contamination levels or typing, collected as part of an epidemiological process, into an inference scheme that targets further investigation on particular food chain operations. For a recent outbreak of staphylococcal food poisoning in Austria, arising from School milk (Schmid et al., 2009), it is clear that a biotracing model, such as the one published by Barker and Gomez-Tome (2011), would have rapidly indicated post process contamination at the dairy as the most likely source of problems. 4.2. Liability issues Although not directly identified with food safety the process of assigning liability is a strong influence of the development of food safety procedures and it is clear that biotracing can influence this 223 process. A mathematical model which connects the transfer of liability with biotracing and indicates a global improvement of food safety as a consequence has been developed by Pouliot and Sumner (2008). In such cases a biotrace acts as a silent witness but in other situations, such as those involving feed and food liability insurance, could support decisions regarding payments. 4.3. Hazard analysis and critical control points(HACCP) Despite the fact that HACCP has been in use for more than 50 years and is recognized as the most effective approach available for producing safe food, it is still a challenge for food companies. More than that, specialists often identify errors at the operational level of many HACCP systems indicating misinterpretation and implementation problems. Critical limits are one of the specific errors made by food companies in their self-monitoring plans designed using the HACCP method (Panunzio, Antoniciello, Pisano, & Rosa, 2007). In such cases, biotracing models may assist food specialists in making calculations for resetting critical limits of microbiological parameters that are crucial for safety of a particular food manufacturing process. 4.4. Risk evaluations/decision making It is clear that biotracing is closely related to risk assessments and, in practice, is driven by similar information supplies and understanding e the hazard domain. However the increased sensitivity of monitors (particularly those based on PCR type techniques) and the increased ability to reason in uncertain situations, that is brought about by systematic application of Bayes theorem and machine learning, ensures that additional decision opportunities arise from a biotracing approach. Decision processes based on risk assessment usually use a threshold for a particular endpoint measure as the trigger for particular action; most visible a shutdown and recall. In contrast a biotracing system might define action levels based on posterior beliefs about sources and hence may pre-empt more intrusive interventions based on endpoints. This preventative or corrective possibility is particularly attractive to decision makers and, potentially, provides an important commercial advantage that can drive the uptake of biotracing. In this respect it is clear that biotracing schemes may also extend the decision process to make use of additional endpoint observations, not necessarily those directly related to food safety, as these may provide evidence that is more easily accessible (i.e. markers) and can support source level inference (Jordan, Wagner, & Hoorfar, 2011). In the future it may become possible to incorporate biotracing at the point of process design to ensure that contamination problems are immediately traceable and unintrusive intervention is possible without safety endpoints moving out of the allowable range. 4.5. Education and training Biotracing can make a significant contribution to education and training for food professionals. Biotracing, uniquely, emphasizes causes rather than effects for food hazards and, therefore, is more directly related to practical considerations and prevention rather than detection and remedial processes. When revealing the cause of a certain microbiological effect that is observed at the endpoint or delivery point of production, students and operators learning about food safety will understand better the importance of correctly executing particular operations to maintain food safety. Nowadays, it is well recognized the importance of education and training in performing correct actions related to food production, 224 A.I. Nicolau et al. / Food Control 29 (2013) 221e225 Fig. 1. Biotracing in action: the endpoint is passed through the core biotrace to discriminate among sources and to detect both critical and operating problems. processing, storage, transport and distribution. Simple prototype implementations, such as SimpleTrace and SimpleMatch, highlight the causal nature involved with biotracing and can act as strong communications tools. 5. Logistic issues for implementation of biotracing The statistical inference that is the core of a biotrace is only one part of a biotracing system that can support improved safety in food chain operations. In practice biotracing involves monitors, alerts, Data from QC and HACCP Data from QC and HACCP Data from QC and HACCP Biotrace 1 Biotrace 2 Biotrace 3 Farm Food Plant Market Safe product Consumer Core biotrace (mathematical model displaying posterior probabilities or likelihood ratios) Operational biotrace (decision taken to assure the food products safety) Fig. 2. Biotracing in food chain configurations; input data streams are linked by biotracing activities. A.I. Nicolau et al. / Food Control 29 (2013) 221e225 data collection, decision support and many other activities. However the biotracing concept remains simple e an output signal is analyzed to discriminate between different components of the input (Fig. 1). Biotracing acts as an accumulator of information and so it is practical for beliefs developed from one tracing operation to behave as inputs to further biotracing; in operational systems different biotracing activities might link along a food chain (Fig. 2). Biotracing is ideally suited to operations that include continuous, in line, data collection. Each new datum can be used automatically to evaluate sources and, in the majority of cases, this will lead to beliefs that no actions are required. However, by using computational techniques usually called ‘machine learning’, the information stream can be used to reduce uncertainties associated with normal chain operations, in so doing, lead to increased confidence in beliefs that indicate abnormal behavior. The learning process can also be configured to account for trends and slow variations, such as seasonality, that sometimes mask significant deviations. 6. Conclusions The case for the introduction of biotracing into food chain management is clear: an effective efficient use of information that helps to eliminate or reduce the uncertainty associated with decisions taken in relation to safety in the food chain. The opportunity for the development has arisen, recently, because of advances in rapid methods for detection and enumeration of organisms and markers and because of significant changes in the collection, storage and manipulation of large volumes of data relating to food manufacture. Increasingly it is apparent that biotracing can also have commercial benefits; after an initial implementation a biotracing system can be operated, alongside existing safety systems, without additional burden and, in practice, may offset substantial costs of chain closure and product recall by indicating targeted preemptive in operation actions. Development of detailed biotracing models includes substantial food chain and microbiological expertise whereas, in other fields, many Bayesian networks are constructed directly from case databases using a computer learning process. As the digital environment for food manufacturing expands the data driven approach to development of biotracing will be increasingly practical. A series of bacterial and physical measurements at strategic time points in a manufacturing process may be used to represent a single ‘case’ and a case database with a few hundred records may be sufficient to capture important correlations and deviations from normal behavior. Machine learning methods may have increasing significance particularly in relation to high throughput data sources that are increasingly relevant to food safety management. Some research units across Europe are already prepared to start tracing pathogens in specific food chains; these include red meat and game, dairy, fish and shellfish. 225 Acknowledgments The work was supported by the European Union-funded Integrated Project BIOTRACER (contract 036272) under the 6th RTD Framework. References Barker, G. C., Gomez, N., & Smid, J. (2009). An introduction to biotracing in food chain systems. Trends in Food Science & Technology, 20(5), 220e226. Barker, G. C., & Gomez-Tome, N. (2011). A risk assessment model for enterotoxigenic Staphylococcus aureus in pasteurized milk: a potential route to source level inference. Risk Analysis, . http://dx.doi.org/10.1111/j.1539-6924.2011.01667.x. Carlos, C., Maretto, D. A., Poppi, R. J., Sato, M. I. Z., & Ottoboni, L. M. M. (2011). Fourier transform infrared microscopy as a bacterial source tracking tool to discriminate fecal E. coli strains. Microchemical Journal, 99, 15e19. Dance, A. (2008). Anthrax case ignites new forensics field. Nature, 454, 813. Foley, S. L., Lynne, A. M., & Nayak, R. (2009). Molecular typing methodologies for microbial source tracking and epidemiological investigations of Gram-negative bacterial foodborne pathogens. Infection, Genetics and Evolution, 9, 430e440. Gorris, L. G. M. (2005). Food safety objective: an integral part of food chain management. Food Control, 16, 801e809. Hoorfar, J., Wagner, M., Jordan, K., Boquin, S. L., & Skiby, J. (2011). Towards biotracing in food chains. International Journal of Food Microbiology, 145(Suppl. 1), S1eS4. http://bbn.ifr.ac.uk/btmodeller/index.php/SimpleTrace (accessed April 2012). http://bbn.ifr.ac.uk/btmodeller/index.php/SimpleMatch (accessed April 2012). http://bbn.ifr.ac.uk (accessed April 2012). http://bbn.ifr.ac.uk/gary/biotracer/agentchain/Chain6.html (accessed April 2012). Jarman, K. H., Kreuzer-Martin, H. W., Wunschel, D. S., Valentine, N. B., Cliff, J. B., Pearson, C. E., et al. (2008). Bayesian-integrated microbial forensics. Applied and Environmental Microbiology, 74, 3573e3582. Jordan, K. N., Wagner, M., & Hoorfar, J. (2011). Biotracing: a new integrated concept in food safety. In J. Hoorfar, K. Jordan, F. Butler, & R. Prugger (Eds.), Food chain integrity e A holistic approach to food traceability, safety, quality and authenticity (pp. 23e37). Cambridge, UK: Woodhead Publishing Limited. Lomonaco, S., Decastelli, L., Nucera, D., Gallina, S., Bianchi, D. M., & Civera, T. (2009). Listeria monocytogenese in Gorgonzola: subtypes, diversity and persistence over time. International Journal of Food Microbiology, 128, 516e520. Panunzio, M. F., Antoniciello, A., Pisano, A., & Rosa, G. (2007). Evaluation of HACCP plans of food industries: case study conducted by the Servizio di Igiene degli Alimenti e della Nutrizione (food and nutrition health service) of the local health authority of Foggia, Italy. International Journal of Environmental Research and Public Health, 4(3), 228e232. Pouliot, S., & Sumner, D. A. (2008). Traceability, liability and incentives for food safety and quality. American Journal of Agricultural Economics, 90, 15e27. Ross, T., & McMeekin, T. A. (2003). Modeling microbial growth within food safety risk assessments. Risk Analysis, 23(1), 179e197. Schmid, D., Fretz, R., Winter, P., Mann, M., Hoger, G., Stoger, A., et al. (2009). Outbreak of staphylococcal food intoxication after consumption of pasteurized milk products. Weiner Klinische Wochencshrift, 121, 125e131. Smid, J. H., Verloo, D., Barker, G. C., & Havelaar, A. H. (2010). Strengths and weaknesses of Monte Carlo simulation models and Bayesian belief networks in microbial risk assessment. International Journal of Food Microbiology, 139(Suppl. 1), S57eS63. Smid, J. H., Swart, A. N., Havelaar, A. H., & Pielaat, A. (2011). A practical framework for the construction of a biotracing model: application to Salmonella in the pork slaughter chain. Risk Analysis, . http://dx.doi.org/10.1111/j.15396924.2011.01591.x. Taroni, F., Aitken, C., Garbolino, P., & Biedermann, A. (2006). Bayesian networks and probabilistic inference in forensic science. Chichester UK: J. Wiley. Verhoef, L., Vennema, H., Van Pelt, W., Lees, D., Boshuizen, H., Henshilwood, K., et al. (2010). Use of norovirus genotype profiles to differentiate origins of food borne outbreak. Emerging Infectious Diseases, 16, 617e624.