Abstract
Introduction
Pediatric adverse drug reactions are difficult to evaluate after approval because pediatric trial evidence is limited and spontaneous-report systems differ in product naming, event coding, seriousness fields, and duplicate submissions.
Methods
We developed a manufacturer-aware benchmark for pediatric pharmacovigilance signal concordance using the FDA Adverse Event Reporting System (FAERS)/AERS as the source system, Canada Vigilance as a cross-system reporting-signal comparator, and JADER for descriptive third-country context. Active ingredients and five clinically adjudicated serious endpoint families were harmonized before detectable manufacturer cross-filing was screened.
Results
The supervised benchmark contained 628 drug–endpoint–period rows: 147 manufacturer-screened Canada Vigilance reporting-signal positives and 481 Canada Vigilance co-report-absent comparators. An endpoint-conditioned FAERS co-reporting comparator achieved an area under the precision–recall curve (AUPRC) of 0.753. Elastic-net logistic regression reached 0.767, and a LightGBM comparator reached 0.774, but clustered paired-difference intervals
versus
the conditioned co-report comparator included zero. In 30 repeated drug-held-out partitions, chemical structure increased the mean AUPRC from 0.764 to 0.796, but the corresponding increment was –0.003 under unseen-scaffold partitions; target and pathway indicators produced no further material increase. Future-period support was concentrated in anaphylaxis: 40 drug pairs remained after manufacturer screening, including 37 with FAERS source exposure, and four of five FAERS scores ranked supported pairs above permutation expectations.
Discussion
Clinical label specificity, detectable duplicate screening, and fair source-signal comparators materially influenced model evaluation. The resulting resource provides a reproducible reference for pediatric signal triage and for testing whether new pharmacological features or models add information beyond established reporting statistics.