All adverse event data comes from the FDA Adverse Event Reporting System (FAERS) through the openFDA API. FAERS is a spontaneous reporting system where healthcare professionals, consumers, and manufacturers submit reports of suspected adverse drug reactions.

Drug labeling data (Boxed Warnings, Contraindications) is pulled in real time from the openFDA Drug Labeling API .

Limitations of FAERS data

• Underreporting: Most adverse events go unreported. No reports does not mean no risk.
• No causation: A report means a patient took a drug and had an event. It does not prove the drug caused it.
• Reporting bias: Media attention, lawsuits, and FDA safety communications can drive spikes in reporting that have nothing to do with true incidence.
• Duplicate reports: The same case can be submitted by the manufacturer, the doctor, and the patient separately.

PRISM uses the Proportional Reporting Ratio (PRR) to measure whether a drug-AE pair is reported more often than expected compared to all other drugs in the FAERS database. PRR is the standard disproportionality metric used by the EMA and was first described by Evans et al. (2001).

Count definitions

PRISM queries four counts from the openFDA API per drug-AE-quarter combination:

Count	Definition
a	Reports with target drug AND target AE
B	All reports with target drug (any AE)
C	All reports with target AE (any drug)
D	All reports in the quarter

B, C, and D are marginal totals (not inner cells of a 2×2 table). Each is obtained from a separate openFDA API query.

This is equivalent to the standard Evans PRR formula [a/(a+b)] / [c/(c+d)] when a is small relative to the marginals, which holds for the vast majority of drug-AE pairs in FAERS. The 95% confidence interval uses the log-normal approximation for ratio measures (Rothman, 2008).

Why PRR and not EBGM or IC?

The FDA uses EBGM (Empirical Bayesian Geometric Mean) internally (DuMouchel, 1999). It applies Bayesian shrinkage to reduce false positives when report counts are low. The WHO Uppsala Monitoring Centre uses the Information Component (IC) (Bate et al., 1998; Norén et al., 2013) for their global VigiBase database.

Both of these methods need access to the full FAERS database to compute the prior distributions that drive the shrinkage. The openFDA API only returns counts for individual queries, not the full reporting distribution. PRR is the right frequentist alternative when working through an API, and it is still the standard at the EMA and MHRA.

A signal is met in a given quarter when all four of the following hold:

Criterion	Threshold	Rationale
Report count (a)	≥ 3	Minimum sample size
PRR	≥ 2.0	Disproportionality
95% CI lower bound	> 1.0	Statistical significance
χ²	≥ 4.0	Independence test

Signal status is based on the most recent 6 quarters:

• CONFIRMED Signal met in 2+ of the last 6 quarters
• EMERGING Signal met in exactly 1 of the last 6 quarters
• NOT DETECTED Signal not met in any of the last 6 quarters

These thresholds come from Evans et al. (2001). We added a CI lower bound > 1 requirement (per Rothman) to reduce false positives in quarters with very few reports.

Signal duration metrics

The Monitor tab reports two complementary duration measures:

• Signal Duration — months since the signal was first detected in any quarter. Used for regulatory timeline comparison against historical lag data.
• Current Streak — consecutive quarters where signal criteria are currently met. Indicates signal persistence and stability.

A long duration with a short streak may indicate an intermittent signal. A short duration with a long streak suggests a newly emerging but consistent signal.

The reference cohort includes 40 drugs across 10 therapeutic classes where FDA took regulatory action (Boxed Warning, Contraindication, Warning, or Withdrawal) after post-market safety signals. These are known, documented cases.

For each drug, we pulled FAERS data for the adverse event that led to the label change, computed PRR per quarter from approval through the label change date, and measured the signal-to-label lag : how long it took from when the FAERS signal first appeared to when FDA acted.

What the cohort tells us

• For some classes (fluoroquinolones, antidiabetics, antithrombotics), FAERS signals showed up well before FDA acted.
• For others (PPIs, bisphosphonates), FAERS was not the driver. FDA acted based on clinical trials and published case series instead.
• Lag times range from under 1 month to over 111 months, so FAERS alone cannot predict when FDA will act.

This is why PRISM shows historical context, not predictions.

• PRISM only uses FAERS data. It does not include clinical trial data, published literature, or international databases like EudraVigilance or VigiBase.
• PRR is a screening tool, not a risk assessment. A signal means something is worth investigating, not that the drug caused the event.
• The openFDA API has a 1 to 3 quarter lag, so the most recent quarter may be incomplete.
• PRISM searches three drug name fields (brand name, generic name, and free-text medicinal product) but can still miss reports with misspellings, abbreviations, or non-US trade names.
• This tool is for educational and research purposes only. It is not regulatory advice.

1. Evans SJW, Waller PC, Davis S. (2001). Use of proportional reporting ratios (PRRs) for signal generation from spontaneous adverse drug reaction reports. Pharmacoepidemiology and Drug Safety , 10(6), 483–486.
2. Rothman KJ, Greenland S, Lash TL. (2008). Modern Epidemiology , 3rd ed. Lippincott Williams & Wilkins. [Log-normal CI for ratio measures]
3. DuMouchel W. (1999). Bayesian data mining in large frequency tables, with an application to the FDA spontaneous reporting system. The American Statistician , 53(3), 177–190. [EBGM/MGPS method used by FDA]
4. Bate A, Lindquist M, Edwards IR, et al. (1998). A Bayesian neural network method for adverse drug reaction signal generation. European Journal of Clinical Pharmacology , 54(4), 315–321. [Original IC method used by WHO/UMC]
5. Norén GN, Hopstadius J, Bate A. (2013). Shrinkage observed-to-expected ratios for robust and transparent large-scale pattern discovery. Statistical Methods in Medical Research , 22(1), 57–69. [Shrinkage IC refinement]
6. FDA. openFDA: FAERS API documentation. https://open.fda.gov/apis/drug/event/
7. FDA. openFDA: Drug Labeling API documentation. https://open.fda.gov/apis/drug/label/