TL;DR: A 14 June 2026 analysis of 28,448 FDA food recalls finds pathogen hazards transfer across companies while labeling and manufacturing compliance signals do not.

A study published 14 June 2026 in Scientific Reports reports that U.S. food recall severity follows two different logics, with implications for how regulators triage risk and allocate attention. According to Scientific Reports, pathogen contamination signals predict the most serious tier across firms, while compliance signals tied to the lowest tier stay firm-specific. The work uses 28,448 openFDA enforcement records covering 2012–2025 to examine structural determinants of severity.

The U.S. Food and Drug Administration classifies food recalls into Class I, Class II and Class III, a decision that drives public notification urgency and regulatory resource allocation, according to Scientific Reports. Class I covers serious adverse health consequences or death, Class II covers temporary or medically reversible effects, and Class III covers unlikely to cause adverse consequences. A separate 2026 compilation of FDA enforcement records published by Fonteum describes the tiers in closely matching terms: Class I as reasonably likely to cause serious adverse health consequences or death, Class II as may cause temporary or medically reversible harm, and Class III as unlikely to cause harm.

Why severity classification strains regulators

The scale described in Scientific Reports is large. Citing DeBeer and colleagues, the paper notes more than 35,000 FDA food recalls between 2002 and 2023, with annual volumes exceeding 3000 in peak years. Each case requires expert evaluation by the Health Hazard Evaluation Committee within two working days. In practice, the paper notes documented delays of months to years between the first reported illness and formal recall announcement in high-profile cases.

The public-health context cited is substantial. According to Scientific Reports, approximately 9.4 million foodborne illnesses, 55,961 hospitalisations, and 1,351 deaths occur annually in the United States, imposing an economic burden exceeding $75 billion per year. The paper places its analysis against the FDA's New Era of Smarter Food Safety blueprint, which identified predictive analytics as a core element of modernised regulatory tools, and against the preventive-control paradigm established by the Food Safety Modernization Act.

A first systematic test on the U.S. food recall repository

Machine learning has been widely tested for food-safety risk assessment, but almost entirely on European data, according to Scientific Reports. On the European Rapid Alert System for Food and Feed database, supervised classifiers achieved accuracies ranging from 74% to 97.8% across various formulations. The authors state that, to date, no peer-reviewed study had attempted systematic severity prediction on the U.S. openFDA food recall database, described as the largest publicly accessible food recall repository in the United States.

The new paper, authored by Juk-Sen Tang, Peilun Li and Junhong Chen, was received 20 March 2026, accepted 10 June 2026 and first online 14 June 2026. The authors declare no competing interests. Funding came from Beijing Academy of Agriculture and Forestry Sciences programs, including the Innovation Capacity Project: Rural Revitalization Research Center and the Reform and Development Program for Agricultural Data Science Technology Research and Application.

Firm history predicts severity because portfolios are stable

Under standard evaluation, described as the Known-firm scenario, a simple baseline assigning each firm its historically most frequent severity class achieves 92% of the best machine learning model's performance. According to Scientific Reports, the Macro-F1 values are 0.818 for the firm-history baseline versus 0.893 for the best machine learning model.

The paper does not treat that pattern as proof that companies inherently cause a given severity. Conditional mutual information analysis finds the concentration is driven primarily by stable hazard-product portfolios within firms. Hazard category mediates 27–63% of the firm–severity association across estimator methods and hazard granularities, with a firm-intrinsic residual of at least 37% under all specifications. The authors describe the result as hazard-portfolio-dominated firm-level patterns with non-negligible firm-intrinsic residual.

For nutrition and food-safety readers, the distinction matters. A company repeatedly associated with one severity tier may largely handle the same kinds of products and hazards over time, rather than carrying a fixed tendency toward serious or minor recalls independent of what it makes. At the same time, the residual of at least 37% leaves room for company-linked factors beyond product mix.

Four regulatory scenarios from familiar firms to first-time recallers

To separate universal from firm-specific drivers, the researchers evaluate predictive models under four regulatory scenarios ranging from triage of known firms to cold-start assessment of first-time recallers. According to Scientific Reports, that design moves from settings where a firm's record is available to settings where the announcing entity has not been seen before.

The design addresses a practical regulatory question: whether a tool tuned on past participants remains useful when a newcomer appears. Food supply chains include established manufacturers, private-label packers, importers and distributors that may enter enforcement data for the first time. A model that depends on recognizing a name will face a different task than a model that recognizes a hazard description.

Pathogen signals travel; compliance signals do not

The central finding is two distinct classes of food-safety signals. Hazard-intrinsic signals, particularly pathogen contamination involving Salmonella and Listeria, are universally associated with Class I severity regardless of the announcing entity. According to Scientific Reports, 92% of pathogen-related Class I recalls were correctly identified even for entirely unseen companies.

That cross-firm transfer was tested against potential cross-announcer coupling. When records with explicit supply-chain markers were excluded, Class I recall changed by only −0.003. The marker filter had a positive predictive value of 76% on manual audit. In other words, the ability to flag serious pathogen recalls did not depend on retaining linked records that might connect related announcers.

In contrast, compliance-related signals, described as labelling defects and GMP violations driving Class III recalls, are almost entirely firm-specific and fail to generalise across entities. A wording pattern that marks a low-severity labeling issue for one company does not reliably mark the same outcome for another company. According to Scientific Reports, machine-learning triage is therefore reliable for pathogen-related recalls but requires mandatory expert review for compliance-related cases involving new entities.

Stricter event-based testing widens the model advantage

The headline SKU-level figures are retained for comparability with prior food-safety machine-learning literature, but the paper adds event-disjoint cross-validation as a stricter deployment-reliability diagnostic. That analysis uses 7,023 distinct contamination events rather than individual stock-keeping-unit records. According to Scientific Reports, the firm-history Macro-F1 is 0.406 and XGBoost is 0.609 in that event-disjoint test.

The increment of XGBoost over firm history expands from +0.075 at SKU level to +0.203 under the stricter unit of analysis, with a 95% confidence interval of [+0.123, +0.283]. The authors present both views in parallel: SKU-level values for comparison with earlier studies, and event-disjoint results as the more demanding indication of how a tool might behave when the same contamination episode generates multiple records.

The paper links that choice to SKU-level pseudoreplication and uncontrolled firm-level autocorrelation. Previously reported machine-learning accuracies of 90% or more on food recall databases likely overstate real-world reliability for those reasons, according to Scientific Reports. The concern is described as relevant to both FDA and EU RASFF research.

Leakage checks and reproducibility floor

Sensitivity analyses examine whether post-classification language in the reason_for_recall field functions as a leakage signal. According to Scientific Reports, such language is present in 0.69% of records and does not function as leakage. Masking effects fall within the cross-version numerical reproducibility floor of absolute change less than 0.03.

That check addresses a common worry in regulatory text modeling: that the narrative written after a severity decision contains the decision itself. The reported result indicates the main patterns do not hinge on that small subset of records. The paper nevertheless calls for entity-overlap disclosure, firm-history baselines and informative null comparisons as minimum standards for future regulatory machine-learning evaluation.

What the results mean for inspection and triage

The authors outline three practical implications. First, firm-level recall history is a powerful risk-profiling tool for targeted regulatory inspections, although its mechanism is hazard-portfolio rather than firm-intrinsic. Second, machine-assisted triage is reliable for pathogen-related recalls but needs mandatory expert review for compliance-related cases involving new entities. Third, earlier high accuracy claims should be interpreted cautiously because overlap between training and testing entities and repeated records from the same event can inflate performance.

For consumers focused on nutrition and healthy living, the actionable distinction is hazard type. Pathogen contamination involving Salmonella and Listeria carries a consistent association with Class I severity across announcing firms in this 2012–2025 enforcement record set. Labeling and good manufacturing practice issues that drive Class III outcomes are more context-dependent and less transferable by automated pattern alone. The study describes plans and recommendations for evaluation standards as recommendations, not as outcomes already adopted.

Sources referenced