Weighted Bayesian network for the classification of unbalanced food safety data: Case study of risk-based monitoring of heavy metals.

Historical data on food safety monitoring often serve as an information source in designing monitoring plans. However, such data are often unbalanced: a small fraction of the dataset refers to food safety hazards that are present in high concentrations (representing commodity batches with a high risk of being contaminated, the positives) and a high fraction of the dataset refers to food safety hazards that are present in low concentrations (representing commodity batches with a low risk of being contaminated, the negatives). Such unbalanced datasets complicate modeling to predict the probability of contamination of commodity batches. This study proposes a weighted Bayesian network (WBN) classifier to improve the model prediction accuracy for the presence of food and feed safety hazards using unbalanced monitoring data, specifically for the presence of heavy metals in feed. Applying different weight values resulted in different classification accuracies for each involved class; the optimal weight value was defined as the value that yielded the most effective monitoring plan, that is, identifying the highest percentage of contaminated feed batches. Results showed that the Bayesian network classifier resulted in a large difference between the classification accuracy of positive samples (20%) and negative samples (99%). With the WBN approach, the classification accuracy of positive samples and negative samples were both around 80%, and the monitoring effectiveness increased from 31% to 80% for pre-set sample size of 3000. Results of this study can be used to improve the effectiveness of monitoring various food safety hazards in food and feed.

Designing optimal food safety monitoring schemes using Bayesian network and integer programming: The case of monitoring dioxins and DL-PCBs.

Efficient food safety monitoring should achieve optimal resource allocation. In this article, a methodology is presented to optimize the use of resources for food safety monitoring aimed at identifying noncompliant samples and estimating background level of hazards in food products. A Bayesian network (BN) model and an optimization model were combined in a single framework. The framework was applied to monitoring dioxins and dioxin-like polychlorinated biphenyls (DL-PCBs) in primary animal-derived food products in the Netherlands. The BN model was built using a national dataset with monitoring results of dioxins and DL-PCBs in animal-derived food products over a 10-year period (2008-2017). These data were used to estimate the probability of detecting suspect samples with dioxins and DL-PCBs levels above preset thresholds, given certain sample conditions. The results of the BN model were then inserted into the optimization model to compute an optimal monitoring scheme. Model estimates showed that the probability of dioxins and DL-PCBs exceeding threshold limits was higher in laying hen eggs and sheep meat than in other animal-derived food (except deer meat). Compared with the monitoring scheme used in the Netherlands in 2018, the optimal monitoring scheme would save around 10,000 EUR per year. This could be obtained by reallocating monitoring resources from products with lower probability of dioxin and DL-PCBs exceeding threshold limits (e.g., pig meat) to products with higher probability (e.g., bovine animal meat), and by shifting sample collection from the last quarter of the year toward the first three quarters of the year.

Cost-Effective Sampling and Analysis for Mycotoxins in a Cereal Batch.

The presence of hazards (e.g., contaminants, pathogens) in food/feed, water, plants, or animals can lead to major economic losses related to human and animal health or the rejection of batches of food or feed. Monitoring these hazards is important but can lead to high costs. This study aimed to find the most cost-effective sampling and analysis (S&A) plan in the cases of the mycotoxins deoxynivalenol (DON) in a wheat batch and aflatoxins (AFB1 ) in a maize batch. An optimization model was constructed, maximizing the number of correct decisions for accepting/rejecting a batch of cereals, with a budget as major constraint. The decision variables were the choice of the analytical method: instrumental method (e.g., liquid chromatography combined with mass-spectrometry (LC-MS/MS)), enzyme-linked-immuno-assay (ELISA), or lateral flow devices (LFD), the number of incremental samples collected from the batch, and the number of aliquots analyzed. S&A plans using ELISA showed to be slightly more cost effective than S&A plans using the other two analytical methods. However, for DON in wheat, the difference between the optimal S&A plans using the three different analytical methods was minimal. For AFB1 in maize, the cost effectiveness of the S&A plan using instrumental methods or ELISA were comparable whereas the S&A plan considering onsite detection with LFDs was least cost effective. In case of nonofficial controls, which do not have to follow official regulations for sampling and analysis, onsite detection with ELISA for both AFB1 in maize and DON in wheat, or with LFDs for DON in wheat, could provide cost-effective alternatives.

Optimization of the Aflatoxin Monitoring Costs along the Maize Supply Chain.

An optimization model was used to gain insight into cost-effective monitoring plans for aflatoxins along the maize supply chain. The model was based on a typical Dutch maize chain, with maize grown in the Black Sea region, and transported by ship to the Netherlands for use as an ingredient in compound feed for dairy cattle. Six different scenarios, with different aflatoxin concentrations at harvest and possible aflatoxin production during transport, were used. By minimizing the costs and using parameters such as the concentration, the variance of the sampling plan, and the monitoring and replacement costs, the model optimized the control points (CPs; e.g., after harvest, before or after transport by sea ship), the number of batches sampled at the CP, and the number of samples per batch. This optimization approach led to an end-of-chain aflatoxin concentration below the predetermined limit. The model showed that, when postharvest aflatoxin production was not possible, it was most cost-effective to collect samples from all batches and replace contaminated batches directly after the harvest, since the replacement costs were the lowest at the origin of the chain. When there was aflatoxin production during storage, it was most cost-effective to collect samples and replace contaminated batches after storage and transport to avoid the duplicate before and after monitoring and replacement costs. Further along the chain a contaminated batch is detected, the more stakeholders are involved, the more expensive the replacement costs and possible recall costs become.

Delaying investments in sensor technology: The rationality of dairy farmers' investment decisions illustrated within the framework of real options theory.

The adoption rate of sensors on dairy farms varies widely. Whereas some sensors are hardly adopted, others are adopted by many farmers. A potential rational explanation for the difference in adoption may be the expected future technological progress in the sensor technology and expected future improved decision support possibilities. For some sensors not much progress can be expected because the technology has already made enormous progress in recent years, whereas for sensors that have only recently been introduced on the market, much progress can be expected. The adoption of sensors may thus be partly explained by uncertainty about the investment decision, in which uncertainty lays in the future performance of the sensors and uncertainty about whether improved informed decision support will become available. The overall aim was to offer a plausible example of why a sensor may not be adopted now. To explain this, the role of uncertainty about technological progress in the investment decision was illustrated for highly adopted sensors (automated estrus detection) and hardly adopted sensors (automated body condition score). This theoretical illustration uses the real options theory, which accounts for the role of uncertainty in the timing of investment decisions. A discrete event model, simulating a farm of 100 dairy cows, was developed to estimate the net present value (NPV) of investing now and investing in 5 yr in both sensor systems. The results show that investing now in automated estrus detection resulted in a higher NPV than investing 5 yr from now, whereas for the automated body condition score postponing the investment resulted in a higher NPV compared with investing now. These results are in line with the observation that farmers postpone investments in sensors. Also, the current high adoption of automated estrus detection sensors can be explained because the NPV of investing now is higher than the NPV of investing in 5 yr. The results confirm that uncertainty about future sensor performance and uncertainty about whether improved decision support will become available play a role in investment decisions.

Effects of incorporating environmental cost and risk aversion on economic values of pig breeding goal traits.

Economic values (EVs) of traits, accounting for environmental impacts and risk preferences of farmers, are required to design breeding goals that contribute to both economic and environmental sustainability. The objective of this study was to assess the effects of incorporating environmental costs and the risk preferences of farmers on the EVs of pig breeding goal traits. A breeding goal consisting of both sow efficiency and production traits was defined for a typical Brazilian farrow-to-finish pig farm with 1,500 productive sows. A mean-variance utility function was employed for deriving the EVs at finishing pig level assuming fixed slaughter weight. The inclusion of risk and risk aversion reduces the economic weights of sow efficiency traits (17%) while increasing the importance of production traits (7%). For a risk-neutral producer, inclusion of environmental cost reduces the economic importance of sow efficiency traits (3%) while increasing the importance of production traits (1%). Genetic changes of breeding goal traits by their genetic standard deviations reduce emissions of greenhouse gases, and excretions of nitrogen and phosphorus per finished pig by up to 6% while increasing farm profit. The estimated EVs could be used to improve selection criteria and thereby contribute to the sustainability of pig production systems.

Systematic Review of Methods to Determine the Cost-Effectiveness of Monitoring Plans for Chemical and Biological Hazards in the Life Sciences.

This study reviews the methods used to determine the cost-effectiveness of monitoring plans for hazards in animals (diseases), plants (pests), soil, water, food, and animal feed, and assesses their applicability to food safety hazards. The review describes the strengths and weaknesses of each method, provides examples of different applications, and concludes with comments about their applicability to food safety. A systematic literature search identified publications assessing the cost-effectiveness of monitoring plans in the life sciences. Publications were classified into 4 groups depending on their subject: food safety, environmental hazards, animal diseases, or pests. Publications were reviewed according to the type of model and input data used, and the types of costs included. Three types of models were used: statistical models, simulation models, and optimization models. Input data were either experimental, historical, or simulated data. Publications differed according to the costs included. More than half the publications only included monitoring costs, whereas other publications included monitoring and management costs, or all costs and benefits. Only a few publications were found in the food safety category and all were relatively recent studies. This suggests that cost-effectiveness analysis of monitoring strategies in food safety is just starting and more research is needed to improve the cost-effectiveness of monitoring hazards in foods.

A conceptual framework for economic optimization of an animal health surveillance portfolio.

Decision making on hazard surveillance in livestock product chains is a multi-hazard, multi-stakeholder, and multi-criteria process that includes a variety of decision alternatives. The multi-hazard aspect means that the allocation of the scarce resource for surveillance should be optimized from the point of view of a surveillance portfolio (SP) rather than a single hazard. In this paper, we present a novel conceptual approach for economic optimization of a SP to address the resource allocation problem for a surveillance organization from a theoretical perspective. This approach uses multi-criteria techniques to evaluate the performances of different settings of a SP, taking cost-benefit aspects of surveillance and stakeholders' preferences into account. The credibility of the approach has also been checked for conceptual validity, data needs and operational validity; the application potentials of the approach are also discussed.

Optimizing bulk milk dioxin monitoring based on costs and effectiveness.

Dioxins are environmental pollutants, potentially present in milk products, which have negative consequences for human health and for the firms and farms involved in the dairy chain. Dioxin monitoring in feed and food has been implemented to detect their presence and estimate their levels in food chains. However, the costs and effectiveness of such programs have not been evaluated. In this study, the costs and effectiveness of bulk milk dioxin monitoring in milk trucks were estimated to optimize the sampling and pooling monitoring strategies aimed at detecting at least 1 contaminated dairy farm out of 20,000 at a target dioxin concentration level. Incidents of different proportions, in terms of the number of contaminated farms, and concentrations were simulated. A combined testing strategy, consisting of screening and confirmatory methods, was assumed as well as testing of pooled samples. Two optimization models were built using linear programming. The first model aimed to minimize monitoring costs subject to a minimum required effectiveness of finding an incident, whereas the second model aimed to maximize the effectiveness for a given monitoring budget. Our results show that a high level of effectiveness is possible, but at high costs. Given specific assumptions, monitoring with 95% effectiveness to detect an incident of 1 contaminated farm at a dioxin concentration of 2 pg of toxic equivalents/g of fat [European Commission's (EC) action level] costs €2.6 million per month. At the same level of effectiveness, a 73% cost reduction is possible when aiming to detect an incident where 2 farms are contaminated at a dioxin concentration of 3 pg of toxic equivalents/g of fat (EC maximum level). With a fixed budget of €40,000 per month, the probability of detecting an incident with a single contaminated farm at a dioxin concentration equal to the EC action level is 4.4%. This probability almost doubled (8.0%) when aiming to detect the same incident but with a dioxin concentration equal to the EC maximum level. This study shows that the effectiveness of finding an incident depends not only on the ratio at which, for testing, collected truck samples are mixed into a pooled sample (aiming at detecting certain concentration), but also the number of collected truck samples. In conclusion, the optimal cost-effective monitoring depends on the number of contaminated farms and the concentration aimed at detection. The models and study results offer quantitative support to risk managers of food industries and food safety authorities.

Comparing technical efficiency of farms with an automatic milking system and a conventional milking system.

Changing from a conventional milking system (CMS) to an automatic milking system (AMS) necessitates a new management approach and a corresponding change in labor tasks. Together with labor savings, AMS farms have been found to have higher capital costs, primarily because of higher maintenance costs and depreciation. Therefore, it is hypothesized that AMS farms differ from CMS farms in capital:labor ratio and possibly their technical efficiency, at least during a transition learning period. The current study used actual farm accounting data from dairy farms in the Netherlands with an AMS and a CMS to investigate the empirical substitution of capital for labor in the AMS farms and to determine if the technical efficiency of the AMS farms differed from the CMS farms. The technical efficiency estimates were obtained with data envelopment analysis. The 63 AMS farms and the 337 CMS farms in the data set did not differ in general farm characteristics such as the number of cows, number of hectares, and the amount of milk quota. Farms with AMS have significantly higher capital costs (€12.71 per 100 kg of milk) than CMS farms (€10.10 per 100 kg of milk). Total labor costs and net outputs were not significantly different between AMS and CMS farms. A clear substitution of capital for labor with the adoption of an AMS could not be observed. Although the AMS farms have a slightly lower technical efficiency (0.76) than the CMS farms (0.78), a significant difference in these estimates was not observed. This indicates that the farms were not different in their ability to use inputs (capital, labor, cows, and land) to produce outputs (total farm revenues). The technical efficiency of farms invested in an AMS in 2008 or earlier was not different from the farms invested in 2009 or 2010, indicating that a learning effect during the transition period was not observed. The results indicate that the economic performance of AMS and CMS farms are similar. What these results show is that other than higher capital costs, the use of AMS rather than a CMS does not affect farm efficiency and that the learning costs to use an AMS are not present as measured by any fall in technical efficiency.

Comparative analysis as a management tool for broiler breeder farms: simulated individual farm analysis (IFAS).

The objective of this study was to develop a management information system to evaluate the tactical management of a breeder flock using individual farm analysis with a deterministic simulation model (IFAS). Individual farm analysis is a method that evaluates the performance of individual farms by comparing them with standards. In the first step of IFAS, a farm accounting system is used to compare performance indicators of a flock with the same performance indicators of the average of a group of flocks that produced in the same time period. In the next step, a deterministic simulation model is used to determine the factors causing the traced deviations in performances. Then, relevant deviations are determined based on the economic and statistical importance of each traced deviation. Finally, the deviations are identified by relevance to give farmers an indication of their strong and weak management practices.

Designing a monitoring program for aflatoxin B1 in feed products using machine learning.

Agricultural commodities used for feed and food production are frequently contaminated with mycotoxins, such as Aflatoxin B1 (AFB1). In Europe, both the government and companies have monitoring programs in place for the presence of AFB1. With limited resources and following risk-based monitoring as prescribed in EU Regulation 2017/625, these monitoring programs focus on batches with the highest probability of being contaminated. This study explored the use of machine learning algorithms (ML) to design risk-based monitoring programs for AFB1 in feed products, considering both monitoring cost and model performance. Historical monitoring data for the presence of AFB1 in feed products (2005-2018; 5605 records in total) were used. Four different ML algorithms, including Decision tree, Logistic regression, Support vector machine and Extreme gradient boosting (XGB), were applied and compared to predict the high-risk feed batches to be considered for further AFB1 sampling and analysis. The monitoring cost included the cost of: sampling and analysis, disease burden, storage, and of recalling and destroying contaminated feed batches. The ML algorithms were able to predict the high-risk batches, with an AUC, recall, and accuracy higher than 0.8, 0.6, and 0.9, respectively. The XGB algorithm outperformed the other three investigated ML. Its incorporation would result into up to 96% reduction in monitoring cost in 2016-2018, as compared to the official monitoring program. The proposed approach for designing risk based monitoring programs can support authorities and industries to reduce the monitoring cost for other food safety hazards as well.

A multi-level hierarchic Markov process with Bayesian updating for herd optimization and simulation in dairy cattle.

Herd optimization models that determine economically optimal insemination and replacement decisions are valuable research tools to study various aspects of farming systems. The aim of this study was to develop a herd optimization and simulation model for dairy cattle. The model determines economically optimal insemination and replacement decisions for individual cows and simulates whole-herd results that follow from optimal decisions. The optimization problem was formulated as a multi-level hierarchic Markov process, and a state space model with Bayesian updating was applied to model variation in milk yield. Methodological developments were incorporated in 2 main aspects. First, we introduced an additional level to the model hierarchy to obtain a more tractable and efficient structure. Second, we included a recently developed cattle feed intake model. In addition to methodological developments, new parameters were used in the state space model and other biological functions. Results were generated for Dutch farming conditions, and outcomes were in line with actual herd performance in the Netherlands. Optimal culling decisions were sensitive to variation in milk yield but insensitive to energy requirements for maintenance and feed intake capacity. We anticipate that the model will be applied in research and extension.

Standardized data in the broiler value chain.

In the Dutch broiler chain, data are collected as a routine practice. However, there is wide variation in the content of data collected and in data collection systems. This variability hampers the use of field data in management information systems to support decisions. The objective of this study was to analyze the quality of data and to standardize the content of data sets in the broiler production chain. To evaluate the quality of data, data sets from 3 Dutch hatcheries, from 23,637 batches of eggs, were assessed. The quality of data was assessed intuitively based on 7 quality attributes. To standardize the content of the data set, a protocol was proposed and validated. The protocol was validated at 30 breeder farms, 3 hatcheries, and 104 broiler farms by using 3 quality attributes: consistency, uniformity, and completeness. Results of the data quality analysis of the 3 Dutch hatcheries showed that the data sets had some fields with inaccurate, incorrect, inconsistent, nonuniform, incomprehensible, missing relevant, or incomplete data. Results of the validation protocol were as follows: feedback was obtained from 23 (77%) breeder farms, 3 (100%) hatcheries, and 7 (7%) broiler farms. Of all the questions, on average 88% were answered on breeder farms; 57, 65, and 82% were answered at each of the 3 hatcheries, respectively; and 79% were answered on the broiler farms. Data collected at 2 hatcheries were more consistent than those collected at the third hatchery. Hatchery data were less consistent than breeder farm data, but the number of data entries at hatcheries far exceeded the number at the farm level. Data from the hatcheries, breeder farms, and broiler farms were not always uniform, possibly because of differences in management strategies. This protocol enables the listing of relevant and standard contents of a data set whereby information exchange along the chain can be simplified. However, it is recommended that the protocol be supplemented with some rules for data collection and management, for example, that variables must be recorded in the provided fields, and that a variable must have one and only one name or code, the same unit of measurement, and the same definition.

Financial losses for Dutch stakeholders during the 2013 aflatoxin incident in Maize in Europe.

Early 2013, high concentrations of aflatoxin M1 were found in the bulk milk of a few dairy farms in the Netherlands. These high concentrations were caused by aflatoxin B1 contaminated maize from Eastern Europe that was processed into compound feed, which was fed to dairy cows. Since the contamination was discovered in the downstream stages of the supply chain, multiple countries and parties were involved and recalls of the feed were necessary, resulting into financial losses. The aim of this study was to estimate the direct short-term financial losses related to the 2013 aflatoxin incident for the maize traders, the feed industry, and the dairy sector in the Netherlands. First, the sequence of events of the incident was retrieved. Then, a Monte Carlo simulation model was built to combine the scarce and uncertain data to estimate the direct financial losses for each stakeholder. The estimated total direct financial losses of this incident were estimated to be between 12 and 25 million euros. The largest share, about 60%, of the total losses was endured by the maize traders. About 39% of the total losses were for the feed industry, and less than 1% of the total losses were for the dairy sector. The financial losses estimated in this study should be interpreted cautiously due to limitations associated with the quality of the data used. Furthermore, this incident led to indirect long-term financial effects, identified but not estimated in this study.

Modeling cost-effective monitoring schemes for food safety contaminants: Case study for dioxins in the dairy supply chain.

Food safety monitoring is essential for hazard identification in food chain, but its application may be limited due to costly analytical methods and (inefficient) sampling procedures. The objective of this study was to design cost-effective monitoring schemes for food safety contaminants along the food production chain, given restricted monitoring budgets. As a case study, we focused on dioxins in the dairy supply chain with feed mills, dairy farms, dairy trucks and storage silos in dairy plants as possible control points. The cost-effectiveness of monitoring schemes was assessed using a model consisting of a simulation module and an optimization module. In the simulation module, the probability to collect at least one contaminated sample was computed for different sampling strategies (simple random sampling, stratified random sampling and systematic sampling) at each control point. The optimization module maximized the effectiveness of a monitoring scheme to identify the contaminated sample by determining the optimal sampling strategies, the optimal number of incremental samples collected, and the pooling rate (number of collected samples mixed into one aggregated sample) at each control point. The modelling approach was applied to two cases with different types of contamination. Results of these cases showed that, to identify the same contaminated sample, monitoring schemes with systematic sampling were more cost-effective at feed mills and dairy farms. The combination of simulation and optimization methods showed to be useful for developing cost-effective food safety monitoring schemes along the food supply chain.

Costs and efficacy of management measures to improve udder health on Dutch dairy farms.

Many different management measures are available to control mastitis, a very costly disease in the dairy sector. The objective of this paper was to evaluate the costs and efficacies of 18 of these management measures, for contagious and environmental pathogens, and their effect on bulk tank somatic cell count (BTSCC) and incidence of clinical mastitis (CM). To determine the efficacies for these management measures, literature data and expertise were combined using Monte Carlo expert evaluation analysis. The effect of management measures varied with the incidence of CM and BTSCC, as well as for environmental and contagious problems. On average, postmilking teat disinfection was found to be the most effective measure in all situations. All management measures had large uncertainty around the most likely value. Results of a data envelopment analysis showed that 4 of the management measures included formed the best-practice frontier (the most cost-efficient measures): keeping cows standing after milking, rinsing clusters after milking a clinical case, using a separate cloth for all cows, and wearing milkers' gloves. Of the top 25 management measures (the 18 base management measures including levels of compliance), 8 were measures with 100% compliance; the others were sublevels of these measures with compliance varying between 25 and 100%. A lower hourly rate of the farmer did not influence management measures from the best-practice frontier, but had some effect on the efficiency scores of the other management measures.

Increasing the revenues from automatic milking by using individual variation in milking characteristics.

The objective of this study was to quantify individual variation in daily milk yield and milking duration in response to the length of the milking interval and to assess the economic potential of using this individual variation to optimize the use of an automated milking system. Random coefficient models were used to describe the individual effects of milking interval on daily milk yield and milking duration. The random coefficient models were fitted on a data set consisting of 4,915 records of normal uninterrupted milkings collected from 311 cows kept in 5 separate herds for 1 wk. The estimated random parameters showed considerable variation between individuals within herds in milk yield and milking duration in response to milking interval. In the actual situation, the herd consisted of 60 cows and the automatic milking system operated at an occupation rate (OR) of 64%. When maximizing daily milk revenues per automated milking system by optimizing individual milking intervals, the average milking interval was reduced from 0.421 d to 0.400 d, the daily milk yield at the herd level was increased from 1,883 to 1,909 kg/d, and milk revenues increased from euro498 to euro507/d. If an OR of 85% could be reached with the same herd size, the optimal milking interval would decrease to 0.238 d, milk yield would increase to 1,997 kg/d, and milk revenues would increase to euro529/d. Consequently, more labor would be required for fetching the cows, and milking duration would increase. Alternatively, an OR of 85% could be achieved by increasing the herd size from 60 to 80 cows without decreasing the milking interval. Milk yield would then increase to 2,535 kg/d and milk revenues would increase to euro673/d. For practical implementation on farms, a dynamic approach is recommended, by which the parameter estimates regarding the effect of interval length on milk yield and the effect of milk yield on milking duration are updated regularly and also the milk production response to concentrate intake is taken into account.

The relation between technical farm performance and antimicrobial use of broiler farms.

The aim of the present study was to explore the relation between both farm performance and antimicrobial use (AMU) of broiler farms. Farm performance was expressed as technical efficiency, obtained by using a bootstrap data envelopment analysis. AMU was expressed as treatment incidence. Cluster analysis is used to obtain groups of farms with similar characteristics regarding technical farm performance and AMU. Results indicate that the farms within the different clusters combine different technical farm performance and different levels of AMU. Between the clusters, significant differences were found in technical farm performance, AMU, the resource intensity of the number of animals at set-up, the number of antimicrobial treatments, the number of antimicrobial treatments related to either gut health or combined problems, and the number of antimicrobial treatments with either yellow or orange active substances. Farmers who combine high levels of AMU with high technical farm performance are likely to overestimate the real economic value of AMU. Proper coordination between the farmer and the veterinarian can be crucial in that case for reducing AMU. Farms with low performance are likely to have poor farm conditions. Improving those farm conditions can help reducing the need for AMU on this kind of farms. The farm-specific conditions have to be considered in future policies aimed at reducing AMU in livestock production.