The new work, described in the Food and Ecological Systems Modelling Journal, re-implements and refines a stochastic risk assessment model originally developed by Perrin and colleagues in 2014. The original version was written in SAS; the updated simulator is written in R and distributed in the Food Safety Knowledge Markup Language, an open format designed to harmonise how risk assessment models are annotated and shared. According to the authors, the model was validated by experts from ANSES, CNIEL, ACTALIA and L2S, and it targets soft mould-ripened cheese produced in France from raw cow milk.

The simulator mirrors the farm-to-fork continuum in a series of hierarchical modules. A farm module first estimates the STEC concentration in the aggregated milk tank that feeds a production batch. Because STEC levels in milk typically fall below the detection limits of routine microbiological methods, the model relies on a relative approach: it measures ordinary E. coli in bulk tank milk and scales it by the ratio of average STEC to average E. coli concentrations in cow faeces, an assumption grounded in the observation that both organisms follow the same faecal routes in dairy herds. The number of cows per farm, the number of infected animals and the bacterial concentrations are all drawn from statistical distributions, including lognormal and Weibull laws, so that natural variability is propagated through every calculation.

Before cheese-making begins, the model can apply a pre-harvest intervention known as milk sorting. Bulk tank milk from each farm is tested for E. coli, and any tanker exceeding a defined threshold is rejected. The frequency of testing is itself a parameter, so producers can explore how often sampling must occur to be effective. The simulator tracks the resulting milk loss and the number of farms discarded, translating food safety decisions into concrete economic terms. With default parameters, a single production batch comprises roughly 22,000 to 23,000 cheeses of 250 grams each, giving the model an industrial scale that matters when weighing costs against risk reduction.

The cheese production module then follows the bacteria through six stages: milk storage, moulding, draining, salting, ripening and cheese storage. Growth during the liquid phase and early solid phase is described by an ordinary differential equation of logistic form, in which the maximum growth rate depends on physico-chemical conditions such as duration, pH, temperature and water activity, following a model published by Augustin and colleagues in 2005. Once milk is transformed into solid cheese, the model shifts perspective: it tracks individual colonies rather than bulk concentration. The number of colonies per cheese is drawn from a Poisson distribution, and each colony grows during draining and salting before entering a decline phase during ripening and storage, with strain-specific inactivation rates.

A crucial refinement distinguishes the five main pathogenic serotypes of STEC from the broader STEC population. The model separates O157 strains from non-O157 pathogenic strains and computes doses for each class independently. In the updated implementation, the proportion of cows carrying the main pathogenic serotypes is applied directly within the farm module, allowing the concentration of pathogenic STEC to be simulated at the source rather than adjusted downstream, a change from the original formulation.

The consumer module converts bacterial numbers into human risk. A serving of 25 grams of cheese is assumed, and the ingested dose combines Poisson-distributed colony counts with lognormal colony sizes that reflect variability between cheeses within a batch. The probability of developing HUS given a dose follows a beta-binomial style relationship in which per-cell risk declines exponentially with age, applied to children from one to fifteen years old, weighted by the proportion of cheese consumed by each age group. The batch risk is obtained by integrating over the dose distribution, and the authors offer two computational routes: straightforward Monte Carlo sampling, or a deterministic quadrature approximation that exploits the monotonic structure of the expectation and converges faster than random sampling.

Post-harvest control takes the form of cheese sampling at the end of ripening, by default on day fourteen of production, adjustable between three and fourteen days. Small portions are tested, and a single positive sample unit rejects the entire batch. Under the model’s assumptions that colonies are homogeneously distributed and that the test detects even a single pathogenic cell, the rejection probability follows directly from the Poisson colony counts. The output module then aggregates many simulated batches to deliver three headline quantities: the overall risk of HUS conditional on a batch reaching the market, the average batch rejection rate, and the average milk loss from sorting.

Several other modifications distinguish the new implementation from its 2014 predecessor. The milk volume attributed to a single cheese is set at 2.2 litres by default rather than 2.5 litres. Intra-cheese variability in colony size is set to zero, meaning all colonies within one cheese share an identical size. Most significantly, the batch risk is now computed at the time of consumption rather than at the end of production, incorporating bacterial inactivation during the storage phase, whose duration follows a triangular distribution updated with more recent parameter values. Because inactivation reduces risk over time, the authors note that averaging over consumption time on a grid slightly overestimates risk, a conservative bias they document explicitly.

The authors are careful about interpretation. They emphasise that outputs such as batch risk, milk loss and rejection rates are estimates from a hypothetical scenario simulated with a state-of-the-art model, not observed prevalence figures, echoing the Food and Agriculture Organization and World Health Organization reminder that models are always incomplete representations of the systems they describe. The real value lies in optimisation: producers and regulators can adjust sorting thresholds, testing frequencies and sampling plans, then observe how relative risk, waste and cost respond. By releasing the model in FSK-ML format, executable within the KNIME workflow platform, the team has turned a specialised French risk assessment into a transparent, reusable tool for anyone working to keep artisanal raw milk cheese both traditional and safe.

Subject of Research: Quantitative microbial risk assessment of Shiga-toxin producing E. coli in raw milk soft cheese

Article Title: Quantitative risk assessment of Haemolytic and Uremic Syndrome (HUS) from consumption of raw milk soft cheese

Article References: Basak, S., Christy, J., Guillier, L., Audiat-Perrin, F., Sanaa, M., Tenenhaus-Aziza, F., Bect, J., & Vazquez, E. (2024). Quantitative risk assessment of Haemolytic and Uremic Syndrome (HUS) from consumption of raw milk soft cheese. Food and Ecological Systems Modelling Journal, 5, Article e109502. https://doi.org/10.3897/fmj.5.109502

Image Credits: AI Generated

DOI: 10.3897/fmj.5.109502

Keywords: STEC, haemolytic uraemic syndrome, raw milk cheese, quantitative microbial risk assessment, food safety, E. coli, risk model, R programming, FSK-ML, cheese production, milk sorting, ArtiSaneFood