Detecting pharmacovigilance signals (disproportionality)

Spontaneous-report databases like the FDA's FAERS are mined for signals of disproportionate reporting (SDR): drug-reaction pairs that occur together more than expected given the background of all reports. The core device is a 2x2 contingency table and a disproportionality metric computed from it — PRR, ROR, EBGM, or IC (BCPNN).

You can build the 2x2 table directly from the public, free OpenFDA /drug/event endpoint (no PHI, no MedDRA license to query; the reaction terms returned are already MedDRA PTs). This skill is statistical screening: a high PRR is a hypothesis, not a confirmed adverse drug reaction.

When to use

• You have a drug of interest and want to see which reactions are over-reported.
• You need a PRR / ROR with confidence interval, or an Empirical Bayes EBGM/EB05 / IC025 to control for the small-count noise PRR/ROR suffer from.
• You are building a routine signal-screening run over OpenFDA or your own aggregated case counts.

The 2x2 table

For one drug D and one reaction R, classify every report:

• PRR = [a/(a+b)] / [c/(c+d)]
• ROR = (a·d)/(b·c)
• IC (BCPNN, log2 information component) ≈ log2( a·(a+b+c+d) / ((a+b)·(a+c)) )
• EBGM = Empirical Bayes Geometric Mean — a gamma-Poisson shrinkage of the observed/expected ratio (the MGPS method) that pulls small-count estimates toward 1; report EB05 (the 5th percentile) as the conservative signal.

Common signal thresholds (screening only): PRR ≥ 2 with χ² ≥ 4 and a ≥ 3; ROR lower 95% CI > 1; IC025 > 0; EB05 ≥ 2.

Quick start (real OpenFDA count queries)

Base endpoint: https://api.fda.gov/drug/event.json. No key needed to try it (240 req/min, 1,000/day per IP; with a free api_key= key: 240/min, 120,000/day). The count=<field>.exact parameter returns a terms histogram, and search= with +AND+ filters the population — that is all you need for a 2x2.

import requests

BASE = "https://api.fda.gov/drug/event.json"

def fda_count(search: str | None, count_field: str) -> int:
    """Total reports matching `search` (sum of the .exact histogram)."""
    params = {"count": count_field}
    if search:
        params["search"] = search
    r = requests.get(BASE, params=params, timeout=30)
    if r.status_code == 404:        # OpenFDA returns 404 for an empty result set
        return 0
    r.raise_for_status()
    return sum(row["count"] for row in r.json()["results"])

def cell_count(search: str | None) -> int:
    """Number of reports matching `search` (use meta.results.total via limit=1)."""
    params = {"limit": 1}
    if search:
        params["search"] = search
    r = requests.get(BASE, params=params, timeout=30)
    if r.status_code == 404:
        return 0
    r.raise_for_status()
    return r.json()["meta"]["results"]["total"]

# Build the 2x2 for warfarin x "gastrointestinal haemorrhage".
DRUG = 'patient.drug.openfda.generic_name:"warfarin"'
RXN  = 'patient.reaction.reactionmeddrapt.exact:"gastrointestinal haemorrhage"'

a = cell_count(f"{DRUG}+AND+{RXN}")          # drug & reaction
b = cell_count(DRUG) - a                      # drug, not reaction
c = cell_count(RXN) - a                       # reaction, not drug
N = cell_count(None)                          # total reports in FAERS
d = N - a - b - c

Compute the metrics from (a, b, c, d):

import math

def prr(a, b, c, d):
    return (a / (a + b)) / (c / (c + d))

def ror(a, b, c, d):
    return (a * d) / (b * c)

def ror_ci(a, b, c, d):
    lnror = math.log((a * d) / (b * c))
    se = math.sqrt(1/a + 1/b + 1/c + 1/d)     # Woolf's method
    lo, hi = math.exp(lnror - 1.96 * se), math.exp(lnror + 1.96 * se)
    return lo, hi

def ic(a, b, c, d):
    n = a + b + c + d
    expected = (a + b) * (a + c) / n
    return math.log2(a / expected) if a and expected else float("nan")

print("PRR", round(prr(a, b, c, d), 2))
print("ROR", round(ror(a, b, c, d), 2), "95% CI", ror_ci(a, b, c, d))
print("IC",  round(ic(a, b, c, d), 2))

For EBGM / EB05 use a maintained Empirical Bayes implementation (e.g. the openEBGM R package or PhViD in R) on the same (a, b, c, d) rather than hand-rolling the gamma-Poisson MGPS shrinkage — the shrinkage prior is the whole point and easy to get wrong.

Workflow

1. Pick the population. Decide your denominator: all of FAERS, or a restricted background (e.g. one drug class, one year via receivedate:[20230101+TO+20231231]). The choice of c/d defines the "expected".
2. Resolve the drug field. Prefer patient.drug.openfda.generic_name (RxNorm ingredient-normalized) over the free-text medicinalproduct to avoid brand fragmentation. Restrict to suspect drugs with patient.drug.drugcharacterization:1 if you want suspect-only signals.
3. Use .exact for the reaction field so "injection site reaction" counts as one phrase, not three words: patient.reaction.reactionmeddrapt.exact.
4. Build the 2x2 with the cell counts above. Verify a + b + c + d == N.
5. Compute PRR and ROR with CIs; add IC025 / EB05 for small counts.
6. Apply thresholds (e.g. PRR ≥ 2, χ² ≥ 4, a ≥ 3) — but treat them as a triage filter, not a verdict.
7. Hand flagged pairs to a safety scientist for medical review, confounder assessment, and labeling/expectedness checks.

Hand-off to / from OpenMed

• From reporting-adverse-events: your own coded, de-identified ICSRs give internal counts you can use instead of or alongside OpenFDA — the same 2x2 math applies. Aggregate only counts; never put narrative PHI in the table.
• From normalizing-rxnorm: normalize the drug name to an RxNorm ingredient before querying so brand/generic synonyms collapse to one cell.
• To querying-openfda-labels: for every signal, check whether the reaction is already on the label (expected) via /drug/label. To reporting-adverse-events: a confirmed signal may require expedited reporting.
• OpenMed runs NER/de-id on-device; only de-identified drug/reaction codes (no PHI) are sent to OpenFDA.

Edge cases & gotchas

• Disproportionality ≠ causality. A high PRR reflects reporting patterns, notoriety bias, and indication confounding — not a proven causal link.
• Small counts break PRR/ROR. With a < 3 the ratios are unstable and CIs explode. This is exactly why EBGM/EB05 and IC025 (shrinkage) exist — prefer them for rare events.
• OpenFDA is a sample, not all of FAERS, and is not deduplicated the way the curated FAERS quarterly files are. Use it for screening; reproduce confirmed signals against the official FAERS extracts.
• .exact is mandatory for counting phrases. Without it, OpenFDA tokenizes the reaction and your counts are wrong.
• OpenFDA returns HTTP 404 for an empty result set (not an empty list) — the helpers above treat 404 as zero. Respect the rate limits; register a free key for routine runs.
• MedDRA versioning. OpenFDA reaction terms are MedDRA PTs at FDA's coding version; if you join to your own MedDRA-coded cases, align the version. MedDRA itself is licensed — you query OpenFDA's already-coded terms, you do not need a MedDRA license to read them, but you do to code your own cases.

Standards & references

• OpenFDA drug adverse event API: https://open.fda.gov/apis/drug/event/
• OpenFDA query syntax (count, .exact, search AND/OR): https://open.fda.gov/apis/query-syntax/
• OpenFDA authentication & rate limits: https://open.fda.gov/apis/authentication/
• Evans et al., PRR for signal generation (Pharmacoepidemiol Drug Saf, 2001): https://pubmed.ncbi.nlm.nih.gov/11828828/
• Bate et al., BCPNN / Information Component (Eur J Clin Pharmacol, 1998): https://pubmed.ncbi.nlm.nih.gov/9696956/
• DuMouchel, Empirical Bayes / MGPS (EBGM): https://www.tandfonline.com/doi/abs/10.1080/00031305.1999.10474456
• CIOMS VIII — Practical Aspects of Signal Detection: https://cioms.ch/publications/product/practical-aspects-of-signal-detection-in-pharmacovigilance/