Signals by drug and event

Note: first load takes a few seconds while the app computes label matches and class stats for the displayed rows. Subsequent interactions are fast.

All (drug, event) pairs flagged by ≥2 of 4 disproportionality methods (GPS/EBGM, PRR, ROR, IC) — ~264k pairs total. The splash shows the top 2000 by Adj EB05 (peak EB05 after Weber-effect shrinkage for drugs <5 years on market); use search to find any other pair. Default filter: Novel pairs with ≥3 quarters of signal. Click any row to see the time-course plot and the label cross-check.

Novel column: “novel” means the event is absent from the drug’s boxed warning, contraindications, warnings, adverse reactions, and indications sections — after MedDRA-synonym expansion (UMLS CUI) and British↔American + clinical-term normalization (anaemia/anemia, adrenocortical/adrenal, lymphoblastic/lymphocytic, etc). Medication- error, product-quality, and administration PTs are hidden. Class co-flags = number of other drugs in the same ATC4 class that also flag this event across the full pair universe (1 = drug-specific; ≥3 suggests class effect). The default view hides pairs with ≥3 class co-flags as likely class effects; clear the Class co-flags column filter to see them. These filters substantially reduce false-positive “novel” flags, but treat remaining rows as hypotheses to investigate, not confirmed novel associations.

Drug-level shows each FAERS-reported drug string as its own row (brand-fragmented: advil, motrin, ibuprofen are separate). Substance-level rolls drugs up to active ingredients via DiAna (86 ibuprofen brands → 1 row), giving stronger per-substance signals. Drugs DiAna can't resolve (~5% of report volume, ~91% of distinct long-tail strings) pass through unchanged in the substance view.
Disclaimer: disproportionate reporting is a statistical pattern, not evidence of causation. Signals are hypotheses requiring further investigation. 'Known' means the event appears in the drug's current FDA label; 'Novel' means it does not (label coverage is limited to drugs we have cached openFDA label data for).

Bayesian Gamma-Poisson Signal Detection

Upload AE report data (CSV with product, event columns) to run disproportionality detection live in the browser. For a richer view with time-stratified signals across historical FAERS data, see the Signals over time tab.

Expected columns: product, event

Each row = one adverse event report.

Signal Detection Results

About faers.mobi

Bayesian and frequentist disproportionality analysis over FAERS (FDA Adverse Event Reporting System) data. Signals are precomputed offline via a deterministic R/targets pipeline and served as read-only parquet for interactive exploration.

Statistical methods

  • GPS — EB05 is the 5th-percentile credible bound of the two-component Gamma mixture posterior (DuMouchel 1999 framework, linear RR scale). This is a direct posterior quantile, not the EBGM geometric mean.
  • PRR + Yates chi-squared (Evans 2001, MHRA criterion)
  • ROR with log-normal CI (van Puijenbroek 2002)
  • BCPNN/IC with 95% credibility bound (Bate 1998, Noren 2006)

Time dimension

Per-quarter rolling 4-quarter window with cumulative-fit prior; EWMA smoothing (lambda = 0.3); no Stan hierarchical model in v1.

Disclaimer

Disproportionate reporting is a statistical pattern, not evidence of causation. Signals are hypotheses requiring further investigation. See FDA FAERS Public Dashboard for official FDA signals.