Toward the next generation of quantitative microbial risk assessment

Quantitative microbial risk assessment (QMRA), as applied to food safety, has had a rich evolution as it enters its fifth decade (Wu and Rodricks, 2020). In their recent Frontiers in Science lead article, Wiedmann et al. (2026) call for the development of a new generation of QMRA modeling that incorporates (i) emerging scientific analytical tools [e.g., genomics, Geographic Information Systems (GIS), agent-based modeling (ABM)], and (ii) a shift toward modeling frameworks that explicitly consider unintended consequences and tradeoffs, using approaches such as multi-criteria decision analysis (MDCA), risk-benefit analysis (RBA), and life-cycle assessment (LCA). Similar calls have been made by others (Taiwo et al., Okoye et al., 2025), with specific reference to predictive microbiology.The tools and approaches highlighted by Wiedmann et al. (2026) these authors share several important commonalities. First, they are systems-focused, aiming to analyze how individual components of a system interact to form a functional unit. This represents a departure from modeling events or phenomena in isolation, the so-called "siloed " approach." Second, whether operating at the micro scale (e.g., microbial signaling) or the macro scale (e.g., LCA to jointly characterize environmental and safety impacts across the lifespan of a food product), these methods are designed to capture dynamic processes, often incorporating temporal and spatial dimensions. Third, the effective use of these tools depends on vast volumes of diverse data. While food safety professionals have touted the value of "big data" for some time, practical applications within QMRA are scant. In short, these newer scientific tools promise larger and more sophisticated mathematical modeling, but that promise has yet to be realized.At its core, QMRA represents an early form of predictive analytics, historically relying on defined datasets and statistical modeling to forecast scenarios that support systematic evaluation of risk and risk-based control strategies. As computing power increased exponentially, predictive analytics evolved to include machine learning (ML), which enables forecasting based on historical data without explicit programming. For instance, ML has been an integral bioinformatics tool, facilitating the analysis of large whole-genome sequencing datasets (Okoye et al., 2025), its application to food microbiology as described by Taiwo et al. (2024). Now a subset of the larger concept of artificial intelligence (AI, i.e., using machines to mimic human thinking), ML has immensealmost unbridled power to handle larger and more complex datasets; identify patterns which may otherwise be missed by statistical analysis alone; interface directly with users; and do this all very quickly, reducing the time to risk-based, data-driven decision-making from months or years, down to days or weeks. In this context, AI serves as an key enabler for integrating emerging food safety tools into QMRA, and for using them in tandem.Interestingly, a search for citations that includes food safety, ML/AI, and QMRA leads to a near dead end, despite frequent discussion of their combined potential. The few existing specific applications are primarily related to genome sequence analysis, where ML-assisted approaches are beginning to inform QRMA. An excellent example is the work of Pouzou et al. (2025), who developed a refined hazard characterization model incorporating genomicallydetermined differences in non-typhoidal Salmonella virulence. The authors used this model in a QMRA to estimate salmonellosis attributable to ground beef. Scenario analysis enabled evaluation of candidate microbiological criteria related to testing, pathogen prevalence, and product diversion. This work is currently informing proposed regulatory changes in U.S. meat and poultry policy. Although not indirectly linked to food safety, Alberts et al. (2024) trained and validated different ML-derived predictive models to identify host species using RNA sequence data from the hemagglutinin gene of H3-type influenza viruses. The authors proposed that this "rapid risk assessment" approach could eventually be adapted to detect species jumps in near real time, with implications for both basic science (strain evolution) and public health (stain emergence). 2024) developed an MCDA-based assessment framework for European Union (EU) food safety incident reports that could, in principle, be incorporated into an AIsupported database for automated risk categorization. The general approach was to identify categories (Reverence, Signal, Reason, and Health Risk), associated subcategories (e.g., related to safety/compliance/public perception, or disease severity), and weighting schemes, which were integrated into a simple ordinal risk categorization algorithm. Hommels et al.(2025) used expert elicitation and a standard MCDA platform to rank key "drivers of change" (delineated as social, technological, economic, environmental, and political) impacting emerging food safety risks. These drivers were categorized in a 2 x 2 matrix based on degree (high or low) of controllability and severity. Interestingly, two specific drivers, environmental contamination and geopolitical conflict, -were identified as both difficult to manage and highly likely to cause severe health impacts, highlighting long-term planning challenges. While neither model has yet to be implemented in AI-driven platforms, the authors articulate a clear vision in which MCDA could be operationalized to support autonomous prediction, early detection, rapid response, and risk prioritization for high-impact food safety issues.At its core, the strength of QMRA lies in its systematic treatment of data and mathematical tools to predict risk. Incorporation of larger datasets, advanced analytic tools, and new platforms should, theoretically, lead to more sophisticated, accurate and accessible models.While this discussion focuses almost exclusively on food safety, for which QMRA is often used to inform governmental actions, its relevance extends to a broad range of stakeholders, including industry scientists and mathematical modelers.Driven by the merging of advanced analytical tools (such as genomics, GIS, MCDA) with ML and AI, the interconnectedness of food safety with public health, environmental sciences, and veterinary medicine (One Heath), as well as sustainability, will become more apparent.In theory, the possibilities are almost limitlesssky is the limit. One can envision ML-or AIdriven visual analytics platforms that allow users to manipulate data and conduct near realtime scenario analyses through web-based interfaces (Chen et al., 2023). In other words, the potential exists for QMRAs to behave like a computer game; think ABM in which the user controls the behavior of the agent(s). These approaches could result in rapid and intricate pattern recognition and prediction, supporting faster responses and more well-informed solutions that consider both costs and benefits. They also hold promise for resource-and riskbased prioritization, potentially establishing a new foundation for science-based risk management.Lest we get ahead of ourselves, it is important to be cautiously optimistic about the future of such advanced QMRA modeling. While more sophisticated models may improve predictive accuracy and responsiveness, consider also the challenges. AI-and ML-based systems can be costly and inaccessible, especially for smaller organizations and stakeholders with limited financial resources. Concerns about data quality, crosstalk between diverse data sets, and data-sharing constraints related to legal or regulatory considerations have been part of the "big data" debate for some time now. Privacy and ethical issues associated with AI are already emerging, while regulatory oversight remains limited. Building these models requires substantial expertise but, perhaps more importantly, collaboration between experts across diverse field soften speaking different technical languages. Rigorous validation will be essential to ensure credibility, accuracy and reliability. Even then, strong technical expertise will be necessary for interpretation of output. These are but a smattering of limitations and constraints, which are further eloquently discussed in Singh et al. (2025).As an educator, I enter this era with some trepidation. To responsibly use these tools, both creators and users must be deeply rooted in the underlying science. Nothing can replace knowledge of the fundamental principles, how they are put into practice, and critical thinking skills, often peppered with real-world experience. We must continue to teach and model these for our young people. Food safety is a "wicked problem." Advanced QMRA tools may give us greater precision and perhaps facilitate faster response, but they don't take away the complexity or the nuances associated with implementing risk management and prioritization in the real world. Human intelligence may be enhanced by a machine, but human sensibility, judgment, and compassion can never be replaced. A child is born in a day and matures over a lifetime. Let us remember to be good parents to the new generation of QMRAs that we are only now conceiving.No generative AI was used in creating any part or section of this manuscript.