Chapter 4

Signal Detection: Quantitative and Qualitative Frameworks

The 2×2 Contingency Table

The foundation of disproportionality analysis

TargetEvent of Interest All OtherOther Events
Drug of Interest ADrug + Event BDrug + Other events
All Other Drugs COther drugs + Event DOther drugs + Other events

Proportional Reporting Ratio (PRR)

PRR Formula

PRR = A / (A + B) C / (C + D)

The proportion of a specific event among all reports for the drug, divided by the proportion of that event among all reports for all other drugs.

Signal of Disproportionate Reporting (SDR) Criteria

PRR ≥ 2
Disproportionality threshold
χ² ≥ 4
Statistical significance
N ≥ 3
Minimum case count

All three criteria must be met simultaneously to flag a signal

Frequentist vs. Bayesian Approaches

Frequentist

PRR

Proportional Reporting Ratio — simple ratio of reporting proportions

Most common metric. Easy to compute but unstable with small cell sizes (few reports).
Bayesian

BCPNN

Bayesian Confidence Propagation Neural Network — used by WHO-UMC

Applies shrinkage algorithms that stabilize results when cell sizes are small. Uses IC (Information Component) as the metric.
Bayesian

MGPS / EBGM

Multi-item Gamma Poisson Shrinker — used by the FDA

Empirical Bayesian Geometric Mean (EBGM) prevents single rare event reports from generating false alarms.

Key Limitations of Spontaneous Reporting

Underreporting

Estimated at 90–99% for some non-serious events. The true incidence of ADRs is substantially higher than what appears in databases.

Weber Effect

Reporting peaks shortly after a drug's launch and then declines — regardless of actual safety profile. Creates a temporal bias in signal detection.

Notoriety Bias

Widely publicized reactions get reported more frequently, inflating their apparent frequency relative to lesser-known ADRs.

Small Cell Instability

With few reports, frequentist methods (PRR) can generate wildly unstable ratios. Bayesian methods with shrinkage are preferred for rare events.

Signal Management Lifecycle

Hover each stage for details

1
Signal Detection
Identify potential signals from statistical screening or case review
→
2
Signal Validation
Determine if the signal represents a genuine safety concern
→
3
Signal Prioritization
Rank signals by clinical impact, strength of evidence, and public health relevance
→
4
Signal Evaluation
In-depth analysis: literature review, clinical trial data, epidemiological assessment
→
5
Recommendation
Label update, DHPC, RMP revision, or further monitoring
→
6
Action & Tracking
Implement regulatory actions and monitor effectiveness

AI-Augmented Signal Detection

Modern signal detection increasingly uses AI and Machine Learning. Algorithms like Random Forests (used in 47% of AI pharmacovigilance studies) can identify statistically significant correlations that human reviewers might miss, particularly in large, high-dimensional datasets.

Pharmacovigilance Manual — Causality Assessment & Signal Detection · Chapter 4