Construction and empirical testing of a verifiable food safety complaint risk-triage model using human foods complaint system data from the United States Food and Drug Administration.

BACKGROUND

Food safety complaint systems contain large numbers of heterogeneous adverse-event reports, but surveillance use requires reproducible data sources and interpretable rules for ranking reports that may involve severe outcomes.

OBJECTIVE

To construct and test a transparent case-level logistic triage model that assigns FDA Human Foods Complaint System (HFCS) reports an urgent-priority score for earlier human review.

METHODS

The official HFCS quarterly workbook was downloaded from the FDA website on 2026-05-13, and case-based reports received from January 1, 2004 through December 31, 2025 were analyzed (n = 149,642). Predictors were derived from validated age and date fields, sex, suspected product descriptors, MedDRA preferred-term counts, symptom-domain indicators, report year, and event-to-receipt lag. Product-group proportions were summarized with unadjusted Wilson 95% confidence intervals and interpreted descriptively. A balanced logistic regression model was assessed on a stratified test set.

RESULTS

Urgent-priority outcomes occurred in 47,093 reports (31.47%). Death appeared in 17,005 reports and hospitalization in 23,405. The hospitalization proportion was 24.9% (95% CI, 24.5-25.3%) among dietary-supplement reports and 18.2% (16.3-20.3%) among infant/baby-food reports; these unadjusted proportions were not interpreted as population risk differences. The model achieved an AUC of 0.723, PR-AUC of 0.526, Brier score of 0.214, Brier skill score of 0.008, a calibration intercept of -0.03, and a calibration slope of 0.91 on 37,411 test cases. Symptom count and the infant/baby-food indicator had the largest positive model weights.

CONCLUSION

HFCS data can support reproducible complaint triage. The model is suitable for ordering reports for human review, whereas product-group contrasts and raw predicted probabilities require cautious interpretation and temporal or external validation.